Remove replaced SIMD implementations

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

@@ -1,8 +1,13 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
+using System.Numerics;
 using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
+using SixLabors.ImageSharp.ColorProfiles;
+using SixLabors.ImageSharp.ColorProfiles.Icc;
 using SixLabors.ImageSharp.Metadata.Profiles.Icc;
 
 namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
@@ -240,6 +245,29 @@ internal abstract partial class JpegColorConverterBase
             IccProfile profile,
             in ComponentValues values,
             float maximumValue)
-            => CmykScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, maximumValue);
+        {
+            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 4);
+            Span<float> packed = memoryOwner.Memory.Span;
+
+            Span<float> c0 = values.Component0;
+            Span<float> c1 = values.Component1;
+            Span<float> c2 = values.Component2;
+            Span<float> c3 = values.Component3;
+
+            // JPEG CMYK stores inverted components, while the ICC converter consumes normalized conventional CMYK.
+            PackedInvertNormalizeInterleave4(c0, c1, c2, c3, packed, maximumValue);
+
+            Span<Cmyk> source = MemoryMarshal.Cast<float, Cmyk>(packed);
+            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed)[..source.Length];
+            ColorConversionOptions options = new()
+            {
+                SourceIccProfile = profile,
+                TargetIccProfile = CompactSrgbV4Profile.Profile,
+            };
+
+            ColorProfileConverter converter = new(options);
+            converter.Convert<Cmyk, Rgb>(source, destination);
+            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
+        }
     }
 }

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

@@ -1,116 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Buffers;
-using System.Numerics;
-using System.Runtime.InteropServices;
-using SixLabors.ImageSharp.ColorProfiles;
-using SixLabors.ImageSharp.ColorProfiles.Icc;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class CmykScalar : JpegColorConverterScalar
-    {
-        public CmykScalar(int precision)
-            : base(JpegColorSpace.Cmyk, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values) =>
-            ConvertToRgbInPlace(values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => ConvertFromRgb(values, this.MaximumValue, rLane, gLane, bLane);
-
-        public static void ConvertToRgbInPlace(in ComponentValues values, float maxValue)
-        {
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-            Span<float> c3 = values.Component3;
-
-            float scale = 1 / (maxValue * maxValue);
-            for (int i = 0; i < c0.Length; i++)
-            {
-                float c = c0[i];
-                float m = c1[i];
-                float y = c2[i];
-                float k = c3[i];
-
-                k *= scale;
-                c0[i] = c * k;
-                c1[i] = m * k;
-                c2[i] = y * k;
-            }
-        }
-
-        public static void ConvertFromRgb(in ComponentValues values, float maxValue, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            Span<float> c = values.Component0;
-            Span<float> m = values.Component1;
-            Span<float> y = values.Component2;
-            Span<float> k = values.Component3;
-
-            for (int i = 0; i < c.Length; i++)
-            {
-                float ctmp = 255f - rLane[i];
-                float mtmp = 255f - gLane[i];
-                float ytmp = 255f - bLane[i];
-                float ktmp = MathF.Min(MathF.Min(ctmp, mtmp), ytmp);
-
-                if (ktmp >= 255f)
-                {
-                    ctmp = 0f;
-                    mtmp = 0f;
-                    ytmp = 0f;
-                }
-                else
-                {
-                    ctmp = (ctmp - ktmp) / (255f - ktmp);
-                    mtmp = (mtmp - ktmp) / (255f - ktmp);
-                    ytmp = (ytmp - ktmp) / (255f - ktmp);
-                }
-
-                c[i] = maxValue - (ctmp * maxValue);
-                m[i] = maxValue - (mtmp * maxValue);
-                y[i] = maxValue - (ytmp * maxValue);
-                k[i] = maxValue - ktmp;
-            }
-        }
-
-        public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maxValue)
-        {
-            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 4);
-            Span<float> packed = memoryOwner.Memory.Span;
-
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-            Span<float> c3 = values.Component3;
-
-            PackedInvertNormalizeInterleave4(c0, c1, c2, c3, packed, maxValue);
-
-            Span<Cmyk> source = MemoryMarshal.Cast<float, Cmyk>(packed);
-            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed)[..source.Length];
-
-            ColorConversionOptions options = new()
-            {
-                SourceIccProfile = profile,
-                TargetIccProfile = CompactSrgbV4Profile.Profile,
-            };
-            ColorProfileConverter converter = new(options);
-            converter.Convert<Cmyk, Rgb>(source, destination);
-
-            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.CmykVector128.cs

@@ -1,100 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class CmykVector128 : JpegColorConverterVector128
-    {
-        public CmykVector128(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
-            Vector128<float> scale = Vector128.Create(1 / (this.MaximumValue * this.MaximumValue));
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint 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 *= scale;
-                c *= k;
-                m *= k;
-                y *= k;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => CmykScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <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));
-
-            Vector128<float> scale = Vector128.Create(maxValue);
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector128<float> ctmp = scale - Unsafe.Add(ref srcR, i);
-                Vector128<float> mtmp = scale - Unsafe.Add(ref srcG, i);
-                Vector128<float> ytmp = scale - Unsafe.Add(ref srcB, i);
-                Vector128<float> ktmp = Vector128.Min(ctmp, Vector128.Min(mtmp, ytmp));
-
-                Vector128<float> kMask = ~Vector128.Equals(ktmp, scale);
-                Vector128<float> divisor = scale - ktmp;
-
-                ctmp = ((ctmp - ktmp) / divisor) & kMask;
-                mtmp = ((mtmp - ktmp) / divisor) & kMask;
-                ytmp = ((ytmp - ktmp) / divisor) & kMask;
-
-                Unsafe.Add(ref destC, i) = scale - (ctmp * scale);
-                Unsafe.Add(ref destM, i) = scale - (mtmp * scale);
-                Unsafe.Add(ref destY, i) = scale - (ytmp * scale);
-                Unsafe.Add(ref destK, i) = scale - ktmp;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.CmykVector256.cs

@@ -1,100 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class CmykVector256 : JpegColorConverterVector256
-    {
-        public CmykVector256(int precision)
-            : base(JpegColorSpace.Cmyk, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-        {
-            ref Vector256<float> c0Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> c1Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> c2Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector256<float> c3Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            // Used for the color conversion
-            Vector256<float> scale = Vector256.Create(1 / (this.MaximumValue * this.MaximumValue));
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector256<float> c = ref Unsafe.Add(ref c0Base, i);
-                ref Vector256<float> m = ref Unsafe.Add(ref c1Base, i);
-                ref Vector256<float> y = ref Unsafe.Add(ref c2Base, i);
-                Vector256<float> k = Unsafe.Add(ref c3Base, i);
-
-                k *= scale;
-                c *= k;
-                m *= k;
-                y *= k;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => CmykScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <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 Vector256<float> destC =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> destM =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> destY =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector256<float> destK =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            ref Vector256<float> srcR =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector256<float> srcG =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector256<float> srcB =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            Vector256<float> scale = Vector256.Create(maxValue);
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector256<float> ctmp = scale - Unsafe.Add(ref srcR, i);
-                Vector256<float> mtmp = scale - Unsafe.Add(ref srcG, i);
-                Vector256<float> ytmp = scale - Unsafe.Add(ref srcB, i);
-                Vector256<float> ktmp = Vector256.Min(ctmp, Vector256.Min(mtmp, ytmp));
-
-                Vector256<float> kMask = ~Vector256.Equals(ktmp, scale);
-                Vector256<float> divisor = scale - ktmp;
-
-                ctmp = ((ctmp - ktmp) / divisor) & kMask;
-                mtmp = ((mtmp - ktmp) / divisor) & kMask;
-                ytmp = ((ytmp - ktmp) / divisor) & kMask;
-
-                Unsafe.Add(ref destC, i) = scale - (ctmp * scale);
-                Unsafe.Add(ref destM, i) = scale - (mtmp * scale);
-                Unsafe.Add(ref destY, i) = scale - (ytmp * scale);
-                Unsafe.Add(ref destK, i) = scale - ktmp;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.CmykVector512.cs

@@ -1,108 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class CmykVector512 : JpegColorConverterVector512
-    {
-        public CmykVector512(int precision)
-            : base(JpegColorSpace.Cmyk, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => CmykScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceVectorized(in ComponentValues values)
-        {
-            ref Vector512<float> c0Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> c1Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> c2Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector512<float> c3Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            // Used for the color conversion
-            Vector512<float> scale = Vector512.Create(1 / (this.MaximumValue * this.MaximumValue));
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector512<float> c = ref Unsafe.Add(ref c0Base, i);
-                ref Vector512<float> m = ref Unsafe.Add(ref c1Base, i);
-                ref Vector512<float> y = ref Unsafe.Add(ref c2Base, i);
-                Vector512<float> k = Unsafe.Add(ref c3Base, i);
-
-                k *= scale;
-                c *= k;
-                m *= k;
-                y *= k;
-            }
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbVectorized(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => ConvertFromRgbVectorized(in values, this.MaximumValue, rLane, gLane, bLane);
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceScalarRemainder(in ComponentValues values)
-             => CmykScalar.ConvertToRgbInPlace(values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbScalarRemainder(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => CmykScalar.ConvertFromRgb(values, this.MaximumValue, rLane, gLane, bLane);
-
-        internal static void ConvertFromRgbVectorized(in ComponentValues values, float maxValue, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector512<float> destC =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> destM =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> destY =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector512<float> destK =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            ref Vector512<float> srcR =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector512<float> srcG =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector512<float> srcB =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            Vector512<float> scale = Vector512.Create(maxValue);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector512<float> ctmp = scale - Unsafe.Add(ref srcR, i);
-                Vector512<float> mtmp = scale - Unsafe.Add(ref srcG, i);
-                Vector512<float> ytmp = scale - Unsafe.Add(ref srcB, i);
-                Vector512<float> ktmp = Vector512.Min(ctmp, Vector512.Min(mtmp, ytmp));
-
-                Vector512<float> kMask = ~Vector512.Equals(ktmp, scale);
-                Vector512<float> divisor = scale - ktmp;
-
-                ctmp = ((ctmp - ktmp) / divisor) & kMask;
-                mtmp = ((mtmp - ktmp) / divisor) & kMask;
-                ytmp = ((ytmp - ktmp) / divisor) & kMask;
-
-                Unsafe.Add(ref destC, i) = scale - (ctmp * scale);
-                Unsafe.Add(ref destM, i) = scale - (mtmp * scale);
-                Unsafe.Add(ref destY, i) = scale - (ytmp * scale);
-                Unsafe.Add(ref destK, i) = scale - ktmp;
-            }
-        }
-    }
-}

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

@@ -1,8 +1,13 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
+using System.Numerics;
 using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
+using SixLabors.ImageSharp.ColorProfiles;
+using SixLabors.ImageSharp.ColorProfiles.Icc;
 using SixLabors.ImageSharp.Common.Helpers;
 using SixLabors.ImageSharp.Metadata.Profiles.Icc;
 
@@ -198,6 +203,28 @@ internal abstract partial class JpegColorConverterBase
             IccProfile profile,
             in ComponentValues values,
             float maximumValue)
-            => GrayScaleScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, maximumValue);
+        {
+            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 3);
+            Span<float> packed = memoryOwner.Memory.Span;
+            Span<float> c0 = values.Component0;
+            Span<float> c1 = values.Component1;
+            Span<float> c2 = values.Component2;
+            float scale = 1F / maximumValue;
+
+            // ICC luminance values are normalized, so the source plane is scaled in place before conversion.
+            TensorPrimitives_.Multiply(c0, scale, c0);
+
+            Span<Y> source = MemoryMarshal.Cast<float, Y>(c0);
+            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed);
+            ColorConversionOptions options = new()
+            {
+                SourceIccProfile = profile,
+                TargetIccProfile = CompactSrgbV4Profile.Profile,
+            };
+
+            ColorProfileConverter converter = new(options);
+            converter.Convert<Y, Rgb>(source, destination);
+            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
+        }
     }
 }

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

@@ -1,97 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Buffers;
-using System.Numerics;
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using SixLabors.ImageSharp.ColorProfiles;
-using SixLabors.ImageSharp.ColorProfiles.Icc;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class GrayScaleScalar : JpegColorConverterScalar
-    {
-        public GrayScaleScalar(int precision)
-            : base(JpegColorSpace.Grayscale, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-            => ConvertToRgbInPlace(in values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => ConvertFromRgbScalar(values, rLane, gLane, bLane);
-
-        internal static void ConvertToRgbInPlace(in ComponentValues values, float maxValue)
-        {
-            ref float c0Base = ref MemoryMarshal.GetReference(values.Component0);
-            ref float c1Base = ref MemoryMarshal.GetReference(values.Component1);
-            ref float c2Base = ref MemoryMarshal.GetReference(values.Component2);
-
-            float scale = 1F / maxValue;
-            for (nuint i = 0; i < (nuint)values.Component0.Length; i++)
-            {
-                float c = Unsafe.Add(ref c0Base, i) * scale;
-
-                Unsafe.Add(ref c0Base, i) = c;
-                Unsafe.Add(ref c1Base, i) = c;
-                Unsafe.Add(ref c2Base, i) = c;
-            }
-        }
-
-        public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maxValue)
-        {
-            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 3);
-            Span<float> packed = memoryOwner.Memory.Span;
-
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-
-            ref float c0Base = ref MemoryMarshal.GetReference(c0);
-            ref float c1Base = ref MemoryMarshal.GetReference(c1);
-            ref float c2Base = ref MemoryMarshal.GetReference(c2);
-
-            float scale = 1F / maxValue;
-            for (nuint i = 0; i < (nuint)values.Component0.Length; i++)
-            {
-                ref float c = ref Unsafe.Add(ref c0Base, i);
-                c *= scale;
-            }
-
-            Span<Y> source = MemoryMarshal.Cast<float, Y>(values.Component0);
-            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed);
-
-            ColorConversionOptions options = new()
-            {
-                SourceIccProfile = profile,
-                TargetIccProfile = CompactSrgbV4Profile.Profile,
-            };
-            ColorProfileConverter converter = new(options);
-            converter.Convert<Y, Rgb>(source, destination);
-
-            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
-        }
-
-        internal static void ConvertFromRgbScalar(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            Span<float> c0 = values.Component0;
-
-            for (int i = 0; i < c0.Length; i++)
-            {
-                // luminosity = (0.299 * r) + (0.587 * g) + (0.114 * b)
-                c0[i] = (float)((0.299f * rLane[i]) + (0.587f * gLane[i]) + (0.114f * bLane[i]));
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.GrayScaleVector128.cs

@@ -1,81 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class GrayScaleVector128 : JpegColorConverterVector128
-    {
-        public GrayScaleVector128(int precision)
-            : base(JpegColorSpace.Grayscale, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => GrayScaleScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <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));
-
-            // Used for the color conversion
-            Vector128<float> scale = Vector128.Create(1 / this.MaximumValue);
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector128<float> c = Unsafe.Add(ref c0Base, i) * scale;
-
-                Unsafe.Add(ref c0Base, i) = c;
-                Unsafe.Add(ref c1Base, i) = c;
-                Unsafe.Add(ref c2Base, i) = c;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector128<float> destLuminance =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component0));
-
-            ref Vector128<float> srcRed =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector128<float> srcGreen =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector128<float> srcBlue =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            // Used for the color conversion
-            Vector128<float> f0299 = Vector128.Create(0.299f);
-            Vector128<float> f0587 = Vector128.Create(0.587f);
-            Vector128<float> f0114 = Vector128.Create(0.114f);
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector128<float> r = ref Unsafe.Add(ref srcRed, i);
-                ref Vector128<float> g = ref Unsafe.Add(ref srcGreen, i);
-                ref Vector128<float> b = ref Unsafe.Add(ref srcBlue, i);
-
-                // luminosity = (0.299 * r) + (0.587 * g) + (0.114 * b)
-                Unsafe.Add(ref destLuminance, i) = Vector128_.MultiplyAddEstimate(f0299, r, Vector128_.MultiplyAddEstimate(f0587, g, f0114 * b));
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.GrayScaleVector256.cs

@@ -1,81 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class GrayScaleVector256 : JpegColorConverterVector256
-    {
-        public GrayScaleVector256(int precision)
-            : base(JpegColorSpace.Grayscale, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-        {
-            ref Vector256<float> c0Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-
-            ref Vector256<float> c1Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-
-            ref Vector256<float> c2Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            // Used for the color conversion
-            Vector256<float> scale = Vector256.Create(1 / this.MaximumValue);
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector256<float> c = Unsafe.Add(ref c0Base, i) * scale;
-
-                Unsafe.Add(ref c0Base, i) = c;
-                Unsafe.Add(ref c1Base, i) = c;
-                Unsafe.Add(ref c2Base, i) = c;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => GrayScaleScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector256<float> destLuminance =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-
-            ref Vector256<float> srcRed =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector256<float> srcGreen =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector256<float> srcBlue =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            // Used for the color conversion
-            Vector256<float> f0299 = Vector256.Create(0.299f);
-            Vector256<float> f0587 = Vector256.Create(0.587f);
-            Vector256<float> f0114 = Vector256.Create(0.114f);
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector256<float> r = ref Unsafe.Add(ref srcRed, i);
-                ref Vector256<float> g = ref Unsafe.Add(ref srcGreen, i);
-                ref Vector256<float> b = ref Unsafe.Add(ref srcBlue, i);
-
-                // luminosity = (0.299 * r) + (0.587 * g) + (0.114 * b)
-                Unsafe.Add(ref destLuminance, i) = Vector256_.MultiplyAddEstimate(f0299, r, Vector256_.MultiplyAddEstimate(f0587, g, f0114 * b));
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.GrayScaleVector512.cs

@@ -1,89 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class GrayScaleVector512 : JpegColorConverterVector512
-    {
-        public GrayScaleVector512(int precision)
-            : base(JpegColorSpace.Grayscale, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => GrayScaleScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceVectorized(in ComponentValues values)
-        {
-            ref Vector512<float> c0Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-
-            ref Vector512<float> c1Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-
-            ref Vector512<float> c2Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            // Used for the color conversion
-            Vector512<float> scale = Vector512.Create(1 / this.MaximumValue);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector512<float> c = Unsafe.Add(ref c0Base, i) * scale;
-
-                Unsafe.Add(ref c0Base, i) = c;
-                Unsafe.Add(ref c1Base, i) = c;
-                Unsafe.Add(ref c2Base, i) = c;
-            }
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbVectorized(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector512<float> destLuminance =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-
-            ref Vector512<float> srcRed =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector512<float> srcGreen =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector512<float> srcBlue =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            // Used for the color conversion
-            Vector512<float> f0299 = Vector512.Create(0.299f);
-            Vector512<float> f0587 = Vector512.Create(0.587f);
-            Vector512<float> f0114 = Vector512.Create(0.114f);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector512<float> r = ref Unsafe.Add(ref srcRed, i);
-                ref Vector512<float> g = ref Unsafe.Add(ref srcGreen, i);
-                ref Vector512<float> b = ref Unsafe.Add(ref srcBlue, i);
-
-                // luminosity = (0.299 * r) + (0.587 * g) + (0.114 * b)
-                Unsafe.Add(ref destLuminance, i) = Vector512_.MultiplyAddEstimate(f0299, r, Vector512_.MultiplyAddEstimate(f0587, g, f0114 * b));
-            }
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceScalarRemainder(in ComponentValues values)
-            => GrayScaleScalar.ConvertToRgbInPlace(in values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbScalarRemainder(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => GrayScaleScalar.ConvertFromRgbScalar(values, rLane, gLane, bLane);
-    }
-}

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

@@ -1,8 +1,13 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
+using System.Numerics;
 using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
+using SixLabors.ImageSharp.ColorProfiles;
+using SixLabors.ImageSharp.ColorProfiles.Icc;
 using SixLabors.ImageSharp.Metadata.Profiles.Icc;
 
 namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
@@ -179,6 +184,26 @@ internal abstract partial class JpegColorConverterBase
             IccProfile profile,
             in ComponentValues values,
             float maximumValue)
-            => RgbScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, maximumValue);
+        {
+            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 3);
+            Span<float> packed = memoryOwner.Memory.Span;
+            Span<float> c0 = values.Component0;
+            Span<float> c1 = values.Component1;
+            Span<float> c2 = values.Component2;
+
+            // JPEG planes use the integer sample domain, while ICC RGB values are normalized and interleaved.
+            PackedNormalizeInterleave3(c0, c1, c2, packed, 1F / maximumValue);
+
+            Span<Rgb> rgb = MemoryMarshal.Cast<float, Rgb>(packed);
+            ColorConversionOptions options = new()
+            {
+                SourceIccProfile = profile,
+                TargetIccProfile = CompactSrgbV4Profile.Profile,
+            };
+
+            ColorProfileConverter converter = new(options);
+            converter.Convert<Rgb, Rgb>(rgb, rgb);
+            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..rgb.Length], c0, c1, c2);
+        }
     }
 }

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

@@ -1,84 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Buffers;
-using System.Numerics;
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using SixLabors.ImageSharp.ColorProfiles;
-using SixLabors.ImageSharp.ColorProfiles.Icc;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class RgbScalar : JpegColorConverterScalar
-    {
-        public RgbScalar(int precision)
-            : base(JpegColorSpace.RGB, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-            => ConvertToRgbInPlace(values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => ConvertFromRgb(values, rLane, gLane, bLane);
-
-        public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maxValue)
-        {
-            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 3);
-            Span<float> packed = memoryOwner.Memory.Span;
-
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-
-            PackedNormalizeInterleave3(c0, c1, c2, packed, 1F / maxValue);
-
-            Span<Rgb> source = MemoryMarshal.Cast<float, Rgb>(packed);
-            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed);
-
-            ColorConversionOptions options = new()
-            {
-                SourceIccProfile = profile,
-                TargetIccProfile = CompactSrgbV4Profile.Profile,
-            };
-            ColorProfileConverter converter = new(options);
-            converter.Convert<Rgb, Rgb>(source, destination);
-
-            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
-        }
-
-        internal static void ConvertToRgbInPlace(ComponentValues values, float maxValue)
-        {
-            ref float c0Base = ref MemoryMarshal.GetReference(values.Component0);
-            ref float c1Base = ref MemoryMarshal.GetReference(values.Component1);
-            ref float c2Base = ref MemoryMarshal.GetReference(values.Component2);
-
-            float scale = 1F / maxValue;
-
-            for (nuint i = 0; i < (nuint)values.Component0.Length; i++)
-            {
-                Unsafe.Add(ref c0Base, i) *= scale;
-                Unsafe.Add(ref c1Base, i) *= scale;
-                Unsafe.Add(ref c2Base, i) *= scale;
-            }
-        }
-
-        internal static void ConvertFromRgb(ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            // TODO: This doesn't seem correct. We should be scaling to the maximum value here.
-            rLane.CopyTo(values.Component0);
-            gLane.CopyTo(values.Component1);
-            bLane.CopyTo(values.Component2);
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.RgbVector128.cs

@@ -1,56 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class RgbVector128 : JpegColorConverterVector128
-    {
-        public RgbVector128(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
-            Vector128<float> scale = Vector128.Create(1 / this.MaximumValue);
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint 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 *= scale;
-                g *= scale;
-                b *= scale;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => RgbScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <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/JpegColorConverter.RgbVector256.cs

@@ -1,56 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class RgbVector256 : JpegColorConverterVector256
-    {
-        public RgbVector256(int precision)
-            : base(JpegColorSpace.RGB, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-        {
-            ref Vector256<float> rBase =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> gBase =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> bBase =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            // Used for the color conversion
-            Vector256<float> scale = Vector256.Create(1 / this.MaximumValue);
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector256<float> r = ref Unsafe.Add(ref rBase, i);
-                ref Vector256<float> g = ref Unsafe.Add(ref gBase, i);
-                ref Vector256<float> b = ref Unsafe.Add(ref bBase, i);
-                r *= scale;
-                g *= scale;
-                b *= scale;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => RgbScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <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/JpegColorConverter.RgbVector512.cs

@@ -1,64 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class RgbVector512 : JpegColorConverterVector512
-    {
-        public RgbVector512(int precision)
-            : base(JpegColorSpace.RGB, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceVectorized(in ComponentValues values)
-        {
-            ref Vector512<float> rBase =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> gBase =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> bBase =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            // Used for the color conversion
-            Vector512<float> scale = Vector512.Create(1 / this.MaximumValue);
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector512<float> r = ref Unsafe.Add(ref rBase, i);
-                ref Vector512<float> g = ref Unsafe.Add(ref gBase, i);
-                ref Vector512<float> b = ref Unsafe.Add(ref bBase, i);
-                r *= scale;
-                g *= scale;
-                b *= scale;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => RgbScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbVectorized(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            rLane.CopyTo(values.Component0);
-            gLane.CopyTo(values.Component1);
-            bLane.CopyTo(values.Component2);
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceScalarRemainder(in ComponentValues values)
-            => RgbScalar.ConvertToRgbInPlace(values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbScalarRemainder(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => RgbScalar.ConvertFromRgb(values, rLane, gLane, bLane);
-    }
-}

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

@@ -1,8 +1,13 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
+using System.Numerics;
 using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
+using SixLabors.ImageSharp.ColorProfiles;
+using SixLabors.ImageSharp.ColorProfiles.Icc;
 using SixLabors.ImageSharp.Metadata.Profiles.Icc;
 
 namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
@@ -154,6 +159,28 @@ internal abstract partial class JpegColorConverterBase
 
         /// <inheritdoc/>
         public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maximumValue)
-            => TiffCmykScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, maximumValue);
+        {
+            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 4);
+            Span<float> packed = memoryOwner.Memory.Span;
+            Span<float> c0 = values.Component0;
+            Span<float> c1 = values.Component1;
+            Span<float> c2 = values.Component2;
+            Span<float> c3 = values.Component3;
+
+            // TIFF CMYK is already non-inverted, so only normalization and interleaving precede ICC conversion.
+            PackedNormalizeInterleave4(c0, c1, c2, c3, packed, maximumValue);
+
+            Span<Cmyk> source = MemoryMarshal.Cast<float, Cmyk>(packed);
+            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed)[..source.Length];
+            ColorConversionOptions options = new()
+            {
+                SourceIccProfile = profile,
+                TargetIccProfile = CompactSrgbV4Profile.Profile,
+            };
+
+            ColorProfileConverter converter = new(options);
+            converter.Convert<Cmyk, Rgb>(source, destination);
+            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
+        }
     }
 }

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

@@ -1,118 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Buffers;
-using System.Numerics;
-using System.Runtime.InteropServices;
-using SixLabors.ImageSharp.ColorProfiles;
-using SixLabors.ImageSharp.ColorProfiles.Icc;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    /// <summary>
-    /// Color converter for tiff images, which use the jpeg compression and CMYK colorspace.
-    /// </summary>
-    internal sealed class TiffCmykScalar : JpegColorConverterScalar
-    {
-        public TiffCmykScalar(int precision)
-            : base(JpegColorSpace.TiffCmyk, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-            => ConvertToRgbInPlace(in values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        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 ConvertToRgbInPlace(in ComponentValues values, float maxValue)
-        {
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-            Span<float> c3 = values.Component3;
-
-            float scale = 1 / maxValue;
-            for (int i = 0; i < c0.Length; i++)
-            {
-                float c = c0[i] * scale;
-                float m = c1[i] * scale;
-                float y = c2[i] * scale;
-                float k = 1 - (c3[i] * scale);
-
-                c0[i] = (1 - c) * k;
-                c1[i] = (1 - m) * k;
-                c2[i] = (1 - y) * k;
-            }
-        }
-
-        public static void ConvertFromRgb(in ComponentValues values, float maxValue, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            Span<float> c = values.Component0;
-            Span<float> m = values.Component1;
-            Span<float> y = values.Component2;
-            Span<float> k = values.Component3;
-
-            for (int i = 0; i < c.Length; i++)
-            {
-                float ctmp = 255F - rLane[i];
-                float mtmp = 255F - gLane[i];
-                float ytmp = 255F - bLane[i];
-                float ktmp = MathF.Min(MathF.Min(ctmp, mtmp), ytmp);
-
-                if (ktmp >= 255F)
-                {
-                    ctmp = 0F;
-                    mtmp = 0F;
-                    ytmp = 0F;
-                }
-                else
-                {
-                    float divisor = 1 / (255F - ktmp);
-                    ctmp = (ctmp - ktmp) * divisor;
-                    mtmp = (mtmp - ktmp) * divisor;
-                    ytmp = (ytmp - ktmp) * divisor;
-                }
-
-                c[i] = ctmp * maxValue;
-                m[i] = mtmp * maxValue;
-                y[i] = ytmp * maxValue;
-                k[i] = ktmp;
-            }
-        }
-
-        public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maxValue)
-        {
-            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 4);
-            Span<float> packed = memoryOwner.Memory.Span;
-
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-            Span<float> c3 = values.Component3;
-
-            PackedNormalizeInterleave4(c0, c1, c2, c3, packed, maxValue);
-
-            Span<Cmyk> source = MemoryMarshal.Cast<float, Cmyk>(packed);
-            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed)[..source.Length];
-
-            ColorConversionOptions options = new()
-            {
-                SourceIccProfile = profile,
-                TargetIccProfile = CompactSrgbV4Profile.Profile,
-            };
-            ColorProfileConverter converter = new(options);
-            converter.Convert<Cmyk, Rgb>(source, destination);
-
-            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.TiffCmykVector128.cs

@@ -1,99 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class TiffCmykVector128 : JpegColorConverterVector128
-    {
-        public TiffCmykVector128(int precision)
-            : base(JpegColorSpace.TiffCmyk, 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));
-
-            Vector128<float> scale = Vector128.Create(1 / this.MaximumValue);
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint 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 = Vector128<float>.One - (k * scale);
-                c = (Vector128<float>.One - (c * scale)) * k;
-                m = (Vector128<float>.One - (m * scale)) * k;
-                y = (Vector128<float>.One - (y * scale)) * k;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => TiffCmykScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <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));
-
-            Vector128<float> scale = Vector128.Create(maxValue);
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector128<float> ctmp = scale - Unsafe.Add(ref srcR, i);
-                Vector128<float> mtmp = scale - Unsafe.Add(ref srcG, i);
-                Vector128<float> ytmp = scale - Unsafe.Add(ref srcB, i);
-                Vector128<float> ktmp = Vector128.Min(ctmp, Vector128.Min(mtmp, ytmp));
-
-                Vector128<float> kMask = ~Vector128.Equals(ktmp, scale);
-                Vector128<float> divisor = Vector128<float>.One / (scale - ktmp);
-
-                ctmp = ((ctmp - ktmp) * divisor) & kMask;
-                mtmp = ((mtmp - ktmp) * divisor) & kMask;
-                ytmp = ((ytmp - ktmp) * divisor) & kMask;
-
-                Unsafe.Add(ref destC, i) = ctmp * scale;
-                Unsafe.Add(ref destM, i) = mtmp * scale;
-                Unsafe.Add(ref destY, i) = ytmp * scale;
-                Unsafe.Add(ref destK, i) = ktmp;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.TiffCmykVector256.cs

@@ -1,99 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class TiffCmykVector256 : JpegColorConverterVector256
-    {
-        public TiffCmykVector256(int precision)
-            : base(JpegColorSpace.TiffCmyk, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-        {
-            ref Vector256<float> c0Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> c1Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> c2Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector256<float> c3Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            Vector256<float> scale = Vector256.Create(1 / this.MaximumValue);
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector256<float> c = ref Unsafe.Add(ref c0Base, i);
-                ref Vector256<float> m = ref Unsafe.Add(ref c1Base, i);
-                ref Vector256<float> y = ref Unsafe.Add(ref c2Base, i);
-                Vector256<float> k = Unsafe.Add(ref c3Base, i);
-
-                k = Vector256<float>.One - (k * scale);
-                c = (Vector256<float>.One - (c * scale)) * k;
-                m = (Vector256<float>.One - (m * scale)) * k;
-                y = (Vector256<float>.One - (y * scale)) * k;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => CmykScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <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 Vector256<float> destC =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> destM =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> destY =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector256<float> destK =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            ref Vector256<float> srcR =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector256<float> srcG =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector256<float> srcB =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            Vector256<float> scale = Vector256.Create(maxValue);
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector256<float> ctmp = scale - Unsafe.Add(ref srcR, i);
-                Vector256<float> mtmp = scale - Unsafe.Add(ref srcG, i);
-                Vector256<float> ytmp = scale - Unsafe.Add(ref srcB, i);
-                Vector256<float> ktmp = Vector256.Min(ctmp, Vector256.Min(mtmp, ytmp));
-
-                Vector256<float> kMask = ~Vector256.Equals(ktmp, scale);
-                Vector256<float> divisor = Vector256<float>.One / (scale - ktmp);
-
-                ctmp = ((ctmp - ktmp) * divisor) & kMask;
-                mtmp = ((mtmp - ktmp) * divisor) & kMask;
-                ytmp = ((ytmp - ktmp) * divisor) & kMask;
-
-                Unsafe.Add(ref destC, i) = ctmp * scale;
-                Unsafe.Add(ref destM, i) = mtmp * scale;
-                Unsafe.Add(ref destY, i) = ytmp * scale;
-                Unsafe.Add(ref destK, i) = ktmp;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.TiffCmykVector512.cs

@@ -1,108 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class TiffCmykVector512 : JpegColorConverterVector512
-    {
-        public TiffCmykVector512(int precision)
-            : base(JpegColorSpace.TiffCmyk, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => TiffCmykScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceVectorized(in ComponentValues values)
-        {
-            ref Vector512<float> c0Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> c1Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> c2Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector512<float> c3Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            // Used for the color conversion
-            Vector512<float> scale = Vector512.Create(1 / this.MaximumValue);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector512<float> c = ref Unsafe.Add(ref c0Base, i);
-                ref Vector512<float> m = ref Unsafe.Add(ref c1Base, i);
-                ref Vector512<float> y = ref Unsafe.Add(ref c2Base, i);
-                Vector512<float> k = Unsafe.Add(ref c3Base, i);
-
-                k = Vector512<float>.One - (k * scale);
-                c = (Vector512<float>.One - (c * scale)) * k;
-                m = (Vector512<float>.One - (m * scale)) * k;
-                y = (Vector512<float>.One - (y * scale)) * k;
-            }
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbVectorized(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => ConvertFromRgbVectorized(in values, this.MaximumValue, rLane, gLane, bLane);
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceScalarRemainder(in ComponentValues values)
-             => TiffCmykScalar.ConvertToRgbInPlace(values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbScalarRemainder(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => TiffCmykScalar.ConvertFromRgb(values, this.MaximumValue, rLane, gLane, bLane);
-
-        internal static void ConvertFromRgbVectorized(in ComponentValues values, float maxValue, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector512<float> destC =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> destM =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> destY =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector512<float> destK =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            ref Vector512<float> srcR =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector512<float> srcG =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector512<float> srcB =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            Vector512<float> scale = Vector512.Create(maxValue);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector512<float> ctmp = scale - Unsafe.Add(ref srcR, i);
-                Vector512<float> mtmp = scale - Unsafe.Add(ref srcG, i);
-                Vector512<float> ytmp = scale - Unsafe.Add(ref srcB, i);
-                Vector512<float> ktmp = Vector512.Min(ctmp, Vector512.Min(mtmp, ytmp));
-
-                Vector512<float> kMask = ~Vector512.Equals(ktmp, scale);
-                Vector512<float> divisor = Vector512<float>.One / (scale - ktmp);
-
-                ctmp = ((ctmp - ktmp) * divisor) & kMask;
-                mtmp = ((mtmp - ktmp) * divisor) & kMask;
-                ytmp = ((ytmp - ktmp) * divisor) & kMask;
-
-                Unsafe.Add(ref destC, i) = ctmp * scale;
-                Unsafe.Add(ref destM, i) = mtmp * scale;
-                Unsafe.Add(ref destY, i) = ytmp * scale;
-                Unsafe.Add(ref destK, i) = ktmp;
-            }
-        }
-    }
-}

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

@@ -1,8 +1,13 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
+using System.Numerics;
 using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
+using SixLabors.ImageSharp.ColorProfiles;
+using SixLabors.ImageSharp.ColorProfiles.Icc;
 using SixLabors.ImageSharp.Common.Helpers;
 using SixLabors.ImageSharp.Metadata.Profiles.Icc;
 
@@ -34,9 +39,9 @@ internal abstract partial class JpegColorConverterBase
 
             // TIFF YccK is non-inverted: decode normalized YCbCr without integer rounding, then let the
             // remaining light after K modulate all three channels.
-            c0 = (y + (YCbCrScalar.RCrMult * cr)) * k;
+            c0 = (y + (YCbCrOperator.RCrMult * cr)) * k;
-            c1 = (y - (YCbCrScalar.GCbMult * cb) - (YCbCrScalar.GCrMult * cr)) * k;
+            c1 = (y - (YCbCrOperator.GCbMult * cb) - (YCbCrOperator.GCrMult * cr)) * k;
-            c2 = (y + (YCbCrScalar.BCbMult * cb)) * k;
+            c2 = (y + (YCbCrOperator.BCbMult * cb)) * k;
         }
 
         /// <inheritdoc/>
@@ -49,9 +54,9 @@ internal abstract partial class JpegColorConverterBase
             Vector128<float> k = Vector128<float>.One - (c3 * scale);
 
             // Four lanes apply the non-rounded YCbCr matrix before their lane-aligned K modulation.
-            c0 = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(YCbCrScalar.RCrMult), y) * k;
+            c0 = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(YCbCrOperator.RCrMult), y) * k;
-            c1 = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(-YCbCrScalar.GCrMult), Vector128_.MultiplyAddEstimate(cb, Vector128.Create(-YCbCrScalar.GCbMult), y)) * k;
+            c1 = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(-YCbCrOperator.GCrMult), Vector128_.MultiplyAddEstimate(cb, Vector128.Create(-YCbCrOperator.GCbMult), y)) * k;
-            c2 = Vector128_.MultiplyAddEstimate(cb, Vector128.Create(YCbCrScalar.BCbMult), y) * k;
+            c2 = Vector128_.MultiplyAddEstimate(cb, Vector128.Create(YCbCrOperator.BCbMult), y) * k;
         }
 
         /// <inheritdoc/>
@@ -64,9 +69,9 @@ internal abstract partial class JpegColorConverterBase
             Vector256<float> k = Vector256<float>.One - (c3 * scale);
 
             // Eight lanes apply the non-rounded YCbCr matrix before their lane-aligned K modulation.
-            c0 = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(YCbCrScalar.RCrMult), y) * k;
+            c0 = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(YCbCrOperator.RCrMult), y) * k;
-            c1 = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(-YCbCrScalar.GCrMult), Vector256_.MultiplyAddEstimate(cb, Vector256.Create(-YCbCrScalar.GCbMult), y)) * k;
+            c1 = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(-YCbCrOperator.GCrMult), Vector256_.MultiplyAddEstimate(cb, Vector256.Create(-YCbCrOperator.GCbMult), y)) * k;
-            c2 = Vector256_.MultiplyAddEstimate(cb, Vector256.Create(YCbCrScalar.BCbMult), y) * k;
+            c2 = Vector256_.MultiplyAddEstimate(cb, Vector256.Create(YCbCrOperator.BCbMult), y) * k;
         }
 
         /// <inheritdoc/>
@@ -79,9 +84,9 @@ internal abstract partial class JpegColorConverterBase
             Vector512<float> k = Vector512<float>.One - (c3 * scale);
 
             // Sixteen lanes apply the non-rounded YCbCr matrix before their lane-aligned K modulation.
-            c0 = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(YCbCrScalar.RCrMult), y) * k;
+            c0 = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(YCbCrOperator.RCrMult), y) * k;
-            c1 = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(-YCbCrScalar.GCrMult), Vector512_.MultiplyAddEstimate(cb, Vector512.Create(-YCbCrScalar.GCbMult), y)) * k;
+            c1 = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(-YCbCrOperator.GCrMult), Vector512_.MultiplyAddEstimate(cb, Vector512.Create(-YCbCrOperator.GCbMult), y)) * k;
-            c2 = Vector512_.MultiplyAddEstimate(cb, Vector512.Create(YCbCrScalar.BCbMult), y) * k;
+            c2 = Vector512_.MultiplyAddEstimate(cb, Vector512.Create(YCbCrOperator.BCbMult), y) * k;
         }
 
         /// <inheritdoc/>
@@ -182,6 +187,31 @@ internal abstract partial class JpegColorConverterBase
 
         /// <inheritdoc/>
         public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maximumValue)
-            => TiffYccKScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, maximumValue);
+        {
+            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 4);
+            Span<float> packed = memoryOwner.Memory.Span;
+            Span<float> c0 = values.Component0;
+            Span<float> c1 = values.Component1;
+            Span<float> c2 = values.Component2;
+            Span<float> c3 = values.Component3;
+
+            // TIFF YccK is non-inverted, so normalize directly before converting its JPEG-specific model to CMYK.
+            PackedNormalizeInterleave4(c0, c1, c2, c3, packed, maximumValue);
+
+            ColorProfileConverter converter = new();
+            Span<Cmyk> source = MemoryMarshal.Cast<float, Cmyk>(packed);
+            converter.Convert<YccK, Cmyk>(MemoryMarshal.Cast<Cmyk, YccK>(source), source);
+
+            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed)[..source.Length];
+            ColorConversionOptions options = new()
+            {
+                SourceIccProfile = profile,
+                TargetIccProfile = CompactSrgbV4Profile.Profile,
+            };
+
+            converter = new ColorProfileConverter(options);
+            converter.Convert<Cmyk, Rgb>(source, destination);
+            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
+        }
     }
 }

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

@@ -1,153 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Buffers;
-using System.Numerics;
-using System.Runtime.InteropServices;
-using SixLabors.ImageSharp.ColorProfiles;
-using SixLabors.ImageSharp.ColorProfiles.Icc;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    /// <summary>
-    /// Color converter for tiff images, which use the jpeg compression and CMYK colorspace.
-    /// </summary>
-    internal sealed class TiffYccKScalar : JpegColorConverterScalar
-    {
-        // Derived from ITU-T Rec. T.871
-        internal const float RCrMult = 1.402f;
-        internal const float GCbMult = (float)(0.114 * 1.772 / 0.587);
-        internal const float GCrMult = (float)(0.299 * 1.402 / 0.587);
-        internal const float BCbMult = 1.772f;
-
-        public TiffYccKScalar(int precision)
-            : base(JpegColorSpace.TiffYccK, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-            => ConvertToRgbInPlace(in values, this.MaximumValue, this.HalfValue);
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => ConvertFromRgb(values, this.HalfValue, this.MaximumValue, rLane, gLane, bLane);
-
-        public static void ConvertToRgbInPlace(in ComponentValues values, float maxValue, float halfValue)
-        {
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-            Span<float> c3 = values.Component3;
-
-            float scale = 1F / maxValue;
-            halfValue *= scale;
-
-            for (int i = 0; i < values.Component0.Length; i++)
-            {
-                float y = c0[i] * scale;
-                float cb = (c1[i] * scale) - halfValue;
-                float cr = (c2[i] * scale) - halfValue;
-                float scaledK = 1 - (c3[i] * scale);
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                c0[i] = (y + (RCrMult * cr)) * scaledK;
-                c1[i] = (y - (GCbMult * cb) - (GCrMult * cr)) * scaledK;
-                c2[i] = (y + (BCbMult * cb)) * scaledK;
-            }
-        }
-
-        public static void ConvertFromRgb(in ComponentValues values, float halfValue, float maxValue, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            Span<float> y = values.Component0;
-            Span<float> cb = values.Component1;
-            Span<float> cr = values.Component2;
-            Span<float> k = values.Component3;
-
-            for (int i = 0; i < cr.Length; i++)
-            {
-                // Scale down to [0-1]
-                const float divisor = 1F / 255F;
-                float r = rLane[i] * divisor;
-                float g = gLane[i] * divisor;
-                float b = bLane[i] * divisor;
-
-                float ytmp;
-                float cbtmp;
-                float crtmp;
-                float ktmp = 1F - MathF.Max(r, MathF.Max(g, b));
-
-                if (ktmp >= 1F)
-                {
-                    ytmp = 0F;
-                    cbtmp = 0.5F;
-                    crtmp = 0.5F;
-                    ktmp = maxValue;
-                }
-                else
-                {
-                    float kmask = 1F / (1F - ktmp);
-                    r *= kmask;
-                    g *= kmask;
-                    b *= kmask;
-
-                    // Scale to [0-maxValue]
-                    ytmp = ((0.299f * r) + (0.587f * g) + (0.114f * b)) * maxValue;
-                    cbtmp = halfValue - (((0.168736f * r) - (0.331264f * g) + (0.5f * b)) * maxValue);
-                    crtmp = halfValue + (((0.5f * r) - (0.418688f * g) - (0.081312f * b)) * maxValue);
-                    ktmp *= maxValue;
-                }
-
-                y[i] = ytmp;
-                cb[i] = cbtmp;
-                cr[i] = crtmp;
-                k[i] = ktmp;
-            }
-        }
-
-        public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maxValue)
-        {
-            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 4);
-            Span<float> packed = memoryOwner.Memory.Span;
-
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-            Span<float> c3 = values.Component3;
-
-            PackedNormalizeInterleave4(c0, c1, c2, c3, packed, maxValue);
-
-            ColorProfileConverter converter = new();
-            Span<Cmyk> source = MemoryMarshal.Cast<float, Cmyk>(packed);
-
-            // YccK is not a defined ICC color space � it's a JPEG-specific encoding used in Adobe-style CMYK JPEGs.
-            // ICC profiles expect colorimetric CMYK values, so we must first convert YccK to CMYK using a hardcoded inverse transform.
-            // This transform assumes Rec.601 YCbCr coefficients and an inverted K channel.
-            //
-            // The YccK => Cmyk conversion is independent of any embedded ICC profile.
-            // Since the same RGB working space is used during conversion to and from XYZ,
-            // colorimetric accuracy is preserved.
-            converter.Convert<YccK, Cmyk>(MemoryMarshal.Cast<Cmyk, YccK>(source), source);
-
-            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed)[..source.Length];
-
-            ColorConversionOptions options = new()
-            {
-                SourceIccProfile = profile,
-                TargetIccProfile = CompactSrgbV4Profile.Profile,
-            };
-            converter = new ColorProfileConverter(options);
-            converter.Convert<Cmyk, Rgb>(source, destination);
-
-            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.TiffYccKVector128.cs

@@ -1,131 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class TiffYccKVector128 : JpegColorConverterVector128
-    {
-        public TiffYccKVector128(int precision)
-            : base(JpegColorSpace.TiffYccK, 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));
-
-            Vector128<float> scale = Vector128.Create(1F / this.MaximumValue);
-            Vector128<float> chromaOffset = Vector128.Create(this.HalfValue) * scale;
-            Vector128<float> rCrMult = Vector128.Create(YCbCrScalar.RCrMult);
-            Vector128<float> gCbMult = Vector128.Create(-YCbCrScalar.GCbMult);
-            Vector128<float> gCrMult = Vector128.Create(-YCbCrScalar.GCrMult);
-            Vector128<float> bCbMult = Vector128.Create(YCbCrScalar.BCbMult);
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector128<float> c0 = ref Unsafe.Add(ref c0Base, i);
-                ref Vector128<float> c1 = ref Unsafe.Add(ref c1Base, i);
-                ref Vector128<float> c2 = ref Unsafe.Add(ref c2Base, i);
-                ref Vector128<float> c3 = ref Unsafe.Add(ref c3Base, i);
-
-                Vector128<float> y = c0 * scale;
-                Vector128<float> cb = (c1 * scale) - chromaOffset;
-                Vector128<float> cr = (c2 * scale) - chromaOffset;
-                Vector128<float> scaledK = Vector128<float>.One - (c3 * scale);
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                Vector128<float> r = Vector128_.MultiplyAddEstimate(cr, rCrMult, y) * scaledK;
-                Vector128<float> g = Vector128_.MultiplyAddEstimate(cr, gCrMult, Vector128_.MultiplyAddEstimate(cb, gCbMult, y)) * scaledK;
-                Vector128<float> b = Vector128_.MultiplyAddEstimate(cb, bCbMult, y) * scaledK;
-
-                c0 = r;
-                c1 = g;
-                c2 = b;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => TiffYccKScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            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));
-
-            ref Vector128<float> destY =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector128<float> destCb =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector128<float> destCr =
-                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));
-
-            Vector128<float> maxSourceValue = Vector128.Create(1 / 255F);
-            Vector128<float> maxSampleValue = Vector128.Create(this.MaximumValue);
-            Vector128<float> chromaOffset = Vector128.Create(this.HalfValue);
-
-            Vector128<float> f0299 = Vector128.Create(0.299f);
-            Vector128<float> f0587 = Vector128.Create(0.587f);
-            Vector128<float> f0114 = Vector128.Create(0.114f);
-            Vector128<float> fn0168736 = Vector128.Create(-0.168736f);
-            Vector128<float> fn0331264 = Vector128.Create(-0.331264f);
-            Vector128<float> fn0418688 = Vector128.Create(-0.418688f);
-            Vector128<float> fn0081312F = Vector128.Create(-0.081312F);
-            Vector128<float> f05 = Vector128.Create(0.5f);
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector128<float> r = Unsafe.Add(ref srcR, i) * maxSourceValue;
-                Vector128<float> g = Unsafe.Add(ref srcG, i) * maxSourceValue;
-                Vector128<float> b = Unsafe.Add(ref srcB, i) * maxSourceValue;
-                Vector128<float> ktmp = Vector128<float>.One - Vector128.Max(r, Vector128.Min(g, b));
-
-                Vector128<float> kMask = ~Vector128.Equals(ktmp, Vector128<float>.One);
-                Vector128<float> divisor = Vector128<float>.One / (Vector128<float>.One - ktmp);
-
-                r = (r * divisor) & kMask;
-                g = (g * divisor) & kMask;
-                b = (b * divisor) & kMask;
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                Vector128<float> y = Vector128_.MultiplyAddEstimate(f0299, r, Vector128_.MultiplyAddEstimate(f0587, g, f0114 * b));
-                Vector128<float> cb = chromaOffset + Vector128_.MultiplyAddEstimate(fn0168736, r, Vector128_.MultiplyAddEstimate(fn0331264, g, f05 * b));
-                Vector128<float> cr = chromaOffset + Vector128_.MultiplyAddEstimate(f05, r, Vector128_.MultiplyAddEstimate(fn0418688, g, fn0081312F * b));
-
-                Unsafe.Add(ref destY, i) = y * maxSampleValue;
-                Unsafe.Add(ref destCb, i) = chromaOffset + (cb * maxSampleValue);
-                Unsafe.Add(ref destCr, i) = chromaOffset + (cr * maxSampleValue);
-                Unsafe.Add(ref destK, i) = ktmp * maxSampleValue;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.TiffYccKVector256.cs

@@ -1,131 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class TiffYccKVector256 : JpegColorConverterVector256
-    {
-        public TiffYccKVector256(int precision)
-            : base(JpegColorSpace.TiffYccK, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-        {
-            ref Vector256<float> c0Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> c1Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> c2Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector256<float> c3Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            Vector256<float> scale = Vector256.Create(1F / this.MaximumValue);
-            Vector256<float> chromaOffset = Vector256.Create(this.HalfValue) * scale;
-            Vector256<float> rCrMult = Vector256.Create(YCbCrScalar.RCrMult);
-            Vector256<float> gCbMult = Vector256.Create(-YCbCrScalar.GCbMult);
-            Vector256<float> gCrMult = Vector256.Create(-YCbCrScalar.GCrMult);
-            Vector256<float> bCbMult = Vector256.Create(YCbCrScalar.BCbMult);
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector256<float> c0 = ref Unsafe.Add(ref c0Base, i);
-                ref Vector256<float> c1 = ref Unsafe.Add(ref c1Base, i);
-                ref Vector256<float> c2 = ref Unsafe.Add(ref c2Base, i);
-                ref Vector256<float> c3 = ref Unsafe.Add(ref c3Base, i);
-
-                Vector256<float> y = c0 * scale;
-                Vector256<float> cb = (c1 * scale) - chromaOffset;
-                Vector256<float> cr = (c2 * scale) - chromaOffset;
-                Vector256<float> scaledK = Vector256<float>.One - (c3 * scale);
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                Vector256<float> r = Vector256_.MultiplyAddEstimate(cr, rCrMult, y) * scaledK;
-                Vector256<float> g = Vector256_.MultiplyAddEstimate(cr, gCrMult, Vector256_.MultiplyAddEstimate(cb, gCbMult, y)) * scaledK;
-                Vector256<float> b = Vector256_.MultiplyAddEstimate(cb, bCbMult, y) * scaledK;
-
-                c0 = r;
-                c1 = g;
-                c2 = b;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => TiffYccKScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector256<float> srcR =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector256<float> srcG =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector256<float> srcB =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            ref Vector256<float> destY =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> destCb =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> destCr =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector256<float> destK =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            Vector256<float> maxSourceValue = Vector256.Create(255F);
-            Vector256<float> maxSampleValue = Vector256.Create(this.MaximumValue);
-            Vector256<float> chromaOffset = Vector256.Create(this.HalfValue);
-
-            Vector256<float> f0299 = Vector256.Create(0.299f);
-            Vector256<float> f0587 = Vector256.Create(0.587f);
-            Vector256<float> f0114 = Vector256.Create(0.114f);
-            Vector256<float> fn0168736 = Vector256.Create(-0.168736f);
-            Vector256<float> fn0331264 = Vector256.Create(-0.331264f);
-            Vector256<float> fn0418688 = Vector256.Create(-0.418688f);
-            Vector256<float> fn0081312F = Vector256.Create(-0.081312F);
-            Vector256<float> f05 = Vector256.Create(0.5f);
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector256<float> r = Unsafe.Add(ref srcR, i) / maxSourceValue;
-                Vector256<float> g = Unsafe.Add(ref srcG, i) / maxSourceValue;
-                Vector256<float> b = Unsafe.Add(ref srcB, i) / maxSourceValue;
-                Vector256<float> ktmp = Vector256<float>.One - Vector256.Max(r, Vector256.Min(g, b));
-
-                Vector256<float> kMask = ~Vector256.Equals(ktmp, Vector256<float>.One);
-                Vector256<float> divisor = Vector256<float>.One / (Vector256<float>.One - ktmp);
-
-                r = (r * divisor) & kMask;
-                g = (g * divisor) & kMask;
-                b = (b * divisor) & kMask;
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                Vector256<float> y = Vector256_.MultiplyAddEstimate(f0299, r, Vector256_.MultiplyAddEstimate(f0587, g, f0114 * b));
-                Vector256<float> cb = chromaOffset + Vector256_.MultiplyAddEstimate(fn0168736, r, Vector256_.MultiplyAddEstimate(fn0331264, g, f05 * b));
-                Vector256<float> cr = chromaOffset + Vector256_.MultiplyAddEstimate(f05, r, Vector256_.MultiplyAddEstimate(fn0418688, g, fn0081312F * b));
-
-                Unsafe.Add(ref destY, i) = y * maxSampleValue;
-                Unsafe.Add(ref destCb, i) = chromaOffset + (cb * maxSampleValue);
-                Unsafe.Add(ref destCr, i) = chromaOffset + (cr * maxSampleValue);
-                Unsafe.Add(ref destK, i) = ktmp * maxSampleValue;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.TiffYccKVector512.cs

@@ -1,142 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class TiffYccKVector512 : JpegColorConverterVector512
-    {
-        public TiffYccKVector512(int precision)
-            : base(JpegColorSpace.TiffYccK, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => TiffYccKScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceVectorized(in ComponentValues values)
-        {
-            ref Vector512<float> c0Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> c1Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> c2Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector512<float> c3Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            Vector512<float> scale = Vector512.Create(1F / this.MaximumValue);
-            Vector512<float> chromaOffset = Vector512.Create(this.HalfValue) * scale;
-            Vector512<float> rCrMult = Vector512.Create(YCbCrScalar.RCrMult);
-            Vector512<float> gCbMult = Vector512.Create(-YCbCrScalar.GCbMult);
-            Vector512<float> gCrMult = Vector512.Create(-YCbCrScalar.GCrMult);
-            Vector512<float> bCbMult = Vector512.Create(YCbCrScalar.BCbMult);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                ref Vector512<float> c0 = ref Unsafe.Add(ref c0Base, i);
-                ref Vector512<float> c1 = ref Unsafe.Add(ref c1Base, i);
-                ref Vector512<float> c2 = ref Unsafe.Add(ref c2Base, i);
-                ref Vector512<float> c3 = ref Unsafe.Add(ref c3Base, i);
-
-                Vector512<float> y = c0 * scale;
-                Vector512<float> cb = (c1 * scale) - chromaOffset;
-                Vector512<float> cr = (c2 * scale) - chromaOffset;
-                Vector512<float> scaledK = Vector512<float>.One - (c3 * scale);
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                Vector512<float> r = Vector512_.MultiplyAddEstimate(cr, rCrMult, y) * scaledK;
-                Vector512<float> g = Vector512_.MultiplyAddEstimate(cr, gCrMult, Vector512_.MultiplyAddEstimate(cb, gCbMult, y)) * scaledK;
-                Vector512<float> b = Vector512_.MultiplyAddEstimate(cb, bCbMult, y) * scaledK;
-
-                c0 = r;
-                c1 = g;
-                c2 = b;
-            }
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbVectorized(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => ConvertFromRgbVectorized(in values, this.MaximumValue, this.HalfValue, rLane, gLane, bLane);
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceScalarRemainder(in ComponentValues values)
-             => TiffYccKScalar.ConvertToRgbInPlace(values, this.MaximumValue, this.HalfValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbScalarRemainder(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => TiffYccKScalar.ConvertFromRgb(values, this.HalfValue, this.MaximumValue, rLane, gLane, bLane);
-
-        internal static void ConvertFromRgbVectorized(in ComponentValues values, float maxValue, float halfValue, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector512<float> srcR =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector512<float> srcG =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector512<float> srcB =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            ref Vector512<float> destY =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> destCb =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> destCr =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector512<float> destK =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            Vector512<float> maxSourceValue = Vector512.Create(255F);
-            Vector512<float> maxSampleValue = Vector512.Create(maxValue);
-            Vector512<float> chromaOffset = Vector512.Create(halfValue);
-
-            Vector512<float> f0299 = Vector512.Create(0.299f);
-            Vector512<float> f0587 = Vector512.Create(0.587f);
-            Vector512<float> f0114 = Vector512.Create(0.114f);
-            Vector512<float> fn0168736 = Vector512.Create(-0.168736f);
-            Vector512<float> fn0331264 = Vector512.Create(-0.331264f);
-            Vector512<float> fn0418688 = Vector512.Create(-0.418688f);
-            Vector512<float> fn0081312F = Vector512.Create(-0.081312F);
-            Vector512<float> f05 = Vector512.Create(0.5f);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector512<float> r = Unsafe.Add(ref srcR, i) / maxSourceValue;
-                Vector512<float> g = Unsafe.Add(ref srcG, i) / maxSourceValue;
-                Vector512<float> b = Unsafe.Add(ref srcB, i) / maxSourceValue;
-                Vector512<float> ktmp = Vector512<float>.One - Vector512.Max(r, Vector512.Min(g, b));
-
-                Vector512<float> kMask = ~Vector512.Equals(ktmp, Vector512<float>.One);
-                Vector512<float> divisor = Vector512<float>.One / (Vector512<float>.One - ktmp);
-
-                r = (r * divisor) & kMask;
-                g = (g * divisor) & kMask;
-                b = (b * divisor) & kMask;
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                Vector512<float> y = Vector512_.MultiplyAddEstimate(f0299, r, Vector512_.MultiplyAddEstimate(f0587, g, f0114 * b));
-                Vector512<float> cb = chromaOffset + Vector512_.MultiplyAddEstimate(fn0168736, r, Vector512_.MultiplyAddEstimate(fn0331264, g, f05 * b));
-                Vector512<float> cr = chromaOffset + Vector512_.MultiplyAddEstimate(f05, r, Vector512_.MultiplyAddEstimate(fn0418688, g, fn0081312F * b));
-
-                Unsafe.Add(ref destY, i) = y * maxSampleValue;
-                Unsafe.Add(ref destCb, i) = chromaOffset + (cb * maxSampleValue);
-                Unsafe.Add(ref destCr, i) = chromaOffset + (cr * maxSampleValue);
-                Unsafe.Add(ref destK, i) = ktmp * maxSampleValue;
-            }
-        }
-    }
-}

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

@@ -1,8 +1,13 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
+using System.Numerics;
 using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
+using SixLabors.ImageSharp.ColorProfiles;
+using SixLabors.ImageSharp.ColorProfiles.Icc;
 using SixLabors.ImageSharp.Common.Helpers;
 using SixLabors.ImageSharp.Metadata.Profiles.Icc;
 
@@ -15,6 +20,26 @@ internal abstract partial class JpegColorConverterBase
     /// </summary>
     internal readonly struct YCbCrOperator : IJpegColorConverterOperator
     {
+        /// <summary>
+        /// The BT.601 red contribution from centered Cr.
+        /// </summary>
+        public const float RCrMult = 1.402F;
+
+        /// <summary>
+        /// The BT.601 green contribution from centered Cb.
+        /// </summary>
+        public const float GCbMult = (float)(0.114 * 1.772 / 0.587);
+
+        /// <summary>
+        /// The BT.601 green contribution from centered Cr.
+        /// </summary>
+        public const float GCrMult = (float)(0.299 * 1.402 / 0.587);
+
+        /// <summary>
+        /// The BT.601 blue contribution from centered Cb.
+        /// </summary>
+        public const float BCbMult = 1.772F;
+
         /// <inheritdoc/>
         public static JpegColorSpace ColorSpace => JpegColorSpace.YCbCr;
 
@@ -40,9 +65,9 @@ internal abstract partial class JpegColorConverterBase
             // the BT.601 matrix, then integer-domain RGB is rounded away from zero and normalized
             // to [nominally] 0..1. Values intentionally remain unclamped because quantizing RGB into the
             // destination pixel format owns saturation; retaining overshoot avoids discarding color information.
-            c0 = MathF.Round(y + (YCbCrScalar.RCrMult * cr), MidpointRounding.AwayFromZero) * scale;
+            c0 = MathF.Round(y + (RCrMult * cr), MidpointRounding.AwayFromZero) * scale;
-            c1 = MathF.Round(y - (YCbCrScalar.GCbMult * cb) - (YCbCrScalar.GCrMult * cr), MidpointRounding.AwayFromZero) * scale;
+            c1 = MathF.Round(y - (GCbMult * cb) - (GCrMult * cr), MidpointRounding.AwayFromZero) * scale;
-            c2 = MathF.Round(y + (YCbCrScalar.BCbMult * cb), MidpointRounding.AwayFromZero) * scale;
+            c2 = MathF.Round(y + (BCbMult * cb), MidpointRounding.AwayFromZero) * scale;
         }
 
         /// <inheritdoc/>
@@ -63,12 +88,12 @@ internal abstract partial class JpegColorConverterBase
             // Lanes are four independent Y/Cb/Cr samples. MultiplyAddEstimate maps to FMA where available:
             // R uses Cr, B uses Cb, and G subtracts both chroma contributions. Rounding occurs in the sample
             // domain before the common normalization scale so all precisions use integer JPEG sample semantics.
-            Vector128<float> r = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(YCbCrScalar.RCrMult), y);
+            Vector128<float> r = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(RCrMult), y);
             Vector128<float> g = Vector128_.MultiplyAddEstimate(
                 cr,
-                Vector128.Create(-YCbCrScalar.GCrMult),
+                Vector128.Create(-GCrMult),
-                Vector128_.MultiplyAddEstimate(cb, Vector128.Create(-YCbCrScalar.GCbMult), y));
+                Vector128_.MultiplyAddEstimate(cb, Vector128.Create(-GCbMult), y));
-            Vector128<float> b = Vector128_.MultiplyAddEstimate(cb, Vector128.Create(YCbCrScalar.BCbMult), y);
+            Vector128<float> b = Vector128_.MultiplyAddEstimate(cb, Vector128.Create(BCbMult), y);
 
             c0 = Vector128_.RoundToNearestInteger(r) * scale;
             c1 = Vector128_.RoundToNearestInteger(g) * scale;
@@ -93,12 +118,12 @@ internal abstract partial class JpegColorConverterBase
             // These eight lanes have the same layout and BT.601 arithmetic as the Vector128 overload.
             // Keeping an explicit overload allows the JIT to emit native YMM operations without a width
             // switch or decomposing the vector into smaller values.
-            Vector256<float> r = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(YCbCrScalar.RCrMult), y);
+            Vector256<float> r = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(RCrMult), y);
             Vector256<float> g = Vector256_.MultiplyAddEstimate(
                 cr,
-                Vector256.Create(-YCbCrScalar.GCrMult),
+                Vector256.Create(-GCrMult),
-                Vector256_.MultiplyAddEstimate(cb, Vector256.Create(-YCbCrScalar.GCbMult), y));
+                Vector256_.MultiplyAddEstimate(cb, Vector256.Create(-GCbMult), y));
-            Vector256<float> b = Vector256_.MultiplyAddEstimate(cb, Vector256.Create(YCbCrScalar.BCbMult), y);
+            Vector256<float> b = Vector256_.MultiplyAddEstimate(cb, Vector256.Create(BCbMult), y);
 
             c0 = Vector256_.RoundToNearestInteger(r) * scale;
             c1 = Vector256_.RoundToNearestInteger(g) * scale;
@@ -123,12 +148,12 @@ internal abstract partial class JpegColorConverterBase
             // Sixteen independent samples occupy the ZMM lanes. The explicit constants are broadcasts;
             // assembly inspection verifies the JIT hoists them from the loop and retains fused operations.
             // The formula and rounding order remain identical to the narrower overloads.
-            Vector512<float> r = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(YCbCrScalar.RCrMult), y);
+            Vector512<float> r = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(RCrMult), y);
             Vector512<float> g = Vector512_.MultiplyAddEstimate(
                 cr,
-                Vector512.Create(-YCbCrScalar.GCrMult),
+                Vector512.Create(-GCrMult),
-                Vector512_.MultiplyAddEstimate(cb, Vector512.Create(-YCbCrScalar.GCbMult), y));
+                Vector512_.MultiplyAddEstimate(cb, Vector512.Create(-GCbMult), y));
-            Vector512<float> b = Vector512_.MultiplyAddEstimate(cb, Vector512.Create(YCbCrScalar.BCbMult), y);
+            Vector512<float> b = Vector512_.MultiplyAddEstimate(cb, Vector512.Create(BCbMult), y);
 
             c0 = Vector512_.RoundToNearestInteger(r) * scale;
             c1 = Vector512_.RoundToNearestInteger(g) * scale;
@@ -258,6 +283,30 @@ internal abstract partial class JpegColorConverterBase
             IccProfile profile,
             in ComponentValues values,
             float maximumValue)
-            => YCbCrScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, maximumValue);
+        {
+            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 3);
+            Span<float> packed = memoryOwner.Memory.Span;
+            Span<float> c0 = values.Component0;
+            Span<float> c1 = values.Component1;
+            Span<float> c2 = values.Component2;
+
+            // ICC profiles rarely expose YCbCr transforms, so BT.601 first produces RGB in the profile's source space.
+            PackedNormalizeInterleave3(c0, c1, c2, packed, 1F / maximumValue);
+
+            ColorProfileConverter converter = new();
+            Span<YCbCr> source = MemoryMarshal.Cast<float, YCbCr>(packed);
+            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed);
+            converter.Convert<YCbCr, Rgb>(source, destination);
+
+            ColorConversionOptions options = new()
+            {
+                SourceIccProfile = profile,
+                TargetIccProfile = CompactSrgbV4Profile.Profile,
+            };
+
+            converter = new ColorProfileConverter(options);
+            converter.Convert<Rgb, Rgb>(destination, destination);
+            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
+        }
     }
 }

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

@@ -1,121 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Buffers;
-using System.Numerics;
-using System.Runtime.InteropServices;
-using SixLabors.ImageSharp.ColorProfiles;
-using SixLabors.ImageSharp.ColorProfiles.Icc;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class YCbCrScalar : JpegColorConverterScalar
-    {
-        // derived from ITU-T Rec. T.871
-        internal const float RCrMult = 1.402f;
-        internal const float GCbMult = (float)(0.114 * 1.772 / 0.587);
-        internal const float GCrMult = (float)(0.299 * 1.402 / 0.587);
-        internal const float BCbMult = 1.772f;
-
-        public YCbCrScalar(int precision)
-            : base(JpegColorSpace.YCbCr, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-            => ConvertToRgbInPlace(values, this.MaximumValue, this.HalfValue);
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => ConvertFromRgb(values, this.HalfValue, rLane, gLane, bLane);
-
-        public static void ConvertToRgbInPlace(in ComponentValues values, float maxValue, float halfValue)
-        {
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-
-            float scale = 1 / maxValue;
-
-            for (int i = 0; i < c0.Length; i++)
-            {
-                float y = c0[i];
-                float cb = c1[i] - halfValue;
-                float cr = c2[i] - halfValue;
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                c0[i] = MathF.Round(y + (RCrMult * cr), MidpointRounding.AwayFromZero) * scale;
-                c1[i] = MathF.Round(y - (GCbMult * cb) - (GCrMult * cr), MidpointRounding.AwayFromZero) * scale;
-                c2[i] = MathF.Round(y + (BCbMult * cb), MidpointRounding.AwayFromZero) * scale;
-            }
-        }
-
-        public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maxValue)
-        {
-            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 3);
-            Span<float> packed = memoryOwner.Memory.Span;
-
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-
-            // Although YCbCr is a defined ICC color space, in practice ICC profiles
-            // do not implement transforms from it.
-            // Therefore, we first convert JPEG YCbCr to RGB manually, then perform
-            // color-managed conversion to the target profile.
-            //
-            // The YCbCr => RGB conversion is based on BT.601 and is independent of any embedded ICC profile.
-            // Since the same RGB working space is used during conversion to and from XYZ,
-            // colorimetric accuracy is preserved.
-            ColorProfileConverter converter = new();
-
-            PackedNormalizeInterleave3(c0, c1, c2, packed, 1F / maxValue);
-
-            Span<YCbCr> source = MemoryMarshal.Cast<float, YCbCr>(packed);
-            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed);
-
-            converter.Convert<YCbCr, Rgb>(source, destination);
-
-            ColorConversionOptions options = new()
-            {
-                SourceIccProfile = profile,
-                TargetIccProfile = CompactSrgbV4Profile.Profile,
-            };
-            converter = new ColorProfileConverter(options);
-            converter.Convert<Rgb, Rgb>(destination, destination);
-
-            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
-        }
-
-        public static void ConvertFromRgb(in ComponentValues values, float halfValue, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            Span<float> y = values.Component0;
-            Span<float> cb = values.Component1;
-            Span<float> cr = values.Component2;
-
-            for (int i = 0; i < y.Length; i++)
-            {
-                float r = rLane[i];
-                float g = gLane[i];
-                float b = bLane[i];
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                y[i] = (0.299f * r) + (0.587f * g) + (0.114f * b);
-                cb[i] = halfValue - (0.168736f * r) - (0.331264f * g) + (0.5f * b);
-                cr[i] = halfValue + (0.5f * r) - (0.418688f * g) - (0.081312f * b);
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.YCbCrVector128.cs

@@ -1,121 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class YCbCrVector128 : JpegColorConverterVector128
-    {
-        public YCbCrVector128(int precision)
-            : base(JpegColorSpace.YCbCr, 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));
-
-            Vector128<float> chromaOffset = Vector128.Create(-this.HalfValue);
-            Vector128<float> scale = Vector128.Create(1 / this.MaximumValue);
-            Vector128<float> rCrMult = Vector128.Create(YCbCrScalar.RCrMult);
-            Vector128<float> gCbMult = Vector128.Create(-YCbCrScalar.GCbMult);
-            Vector128<float> gCrMult = Vector128.Create(-YCbCrScalar.GCrMult);
-            Vector128<float> bCbMult = Vector128.Create(YCbCrScalar.BCbMult);
-
-            // Walking 8 elements at one step:
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                // y = yVals[i];
-                // cb = cbVals[i] - 128F;
-                // cr = crVals[i] - 128F;
-                ref Vector128<float> c0 = ref Unsafe.Add(ref c0Base, i);
-                ref Vector128<float> c1 = ref Unsafe.Add(ref c1Base, i);
-                ref Vector128<float> c2 = ref Unsafe.Add(ref c2Base, i);
-
-                Vector128<float> y = c0;
-                Vector128<float> cb = c1 + chromaOffset;
-                Vector128<float> cr = c2 + chromaOffset;
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                Vector128<float> r = Vector128_.MultiplyAddEstimate(cr, rCrMult, y);
-                Vector128<float> g = Vector128_.MultiplyAddEstimate(cr, gCrMult, Vector128_.MultiplyAddEstimate(cb, gCbMult, y));
-                Vector128<float> b = Vector128_.MultiplyAddEstimate(cb, bCbMult, y);
-
-                r = Vector128_.RoundToNearestInteger(r) * scale;
-                g = Vector128_.RoundToNearestInteger(g) * scale;
-                b = Vector128_.RoundToNearestInteger(b) * scale;
-
-                c0 = r;
-                c1 = g;
-                c2 = b;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => YCbCrScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector128<float> destY =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector128<float> destCb =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector128<float> destCr =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            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));
-
-            Vector128<float> chromaOffset = Vector128.Create(this.HalfValue);
-            Vector128<float> f0299 = Vector128.Create(0.299f);
-            Vector128<float> f0587 = Vector128.Create(0.587f);
-            Vector128<float> f0114 = Vector128.Create(0.114f);
-            Vector128<float> fn0168736 = Vector128.Create(-0.168736f);
-            Vector128<float> fn0331264 = Vector128.Create(-0.331264f);
-            Vector128<float> fn0418688 = Vector128.Create(-0.418688f);
-            Vector128<float> fn0081312F = Vector128.Create(-0.081312F);
-            Vector128<float> f05 = Vector128.Create(0.5f);
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector128<float> r = Unsafe.Add(ref srcR, i);
-                Vector128<float> g = Unsafe.Add(ref srcG, i);
-                Vector128<float> b = Unsafe.Add(ref srcB, i);
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                Vector128<float> y = Vector128_.MultiplyAddEstimate(f0299, r, Vector128_.MultiplyAddEstimate(f0587, g, f0114 * b));
-                Vector128<float> cb = chromaOffset + Vector128_.MultiplyAddEstimate(fn0168736, r, Vector128_.MultiplyAddEstimate(fn0331264, g, f05 * b));
-                Vector128<float> cr = chromaOffset + Vector128_.MultiplyAddEstimate(f05, r, Vector128_.MultiplyAddEstimate(fn0418688, g, fn0081312F * b));
-
-                Unsafe.Add(ref destY, i) = y;
-                Unsafe.Add(ref destCb, i) = cb;
-                Unsafe.Add(ref destCr, i) = cr;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.YCbCrVector256.cs

@@ -1,121 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class YCbCrVector256 : JpegColorConverterVector256
-    {
-        public YCbCrVector256(int precision)
-            : base(JpegColorSpace.YCbCr, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-        {
-            ref Vector256<float> c0Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> c1Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> c2Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            Vector256<float> chromaOffset = Vector256.Create(-this.HalfValue);
-            Vector256<float> scale = Vector256.Create(1 / this.MaximumValue);
-            Vector256<float> rCrMult = Vector256.Create(YCbCrScalar.RCrMult);
-            Vector256<float> gCbMult = Vector256.Create(-YCbCrScalar.GCbMult);
-            Vector256<float> gCrMult = Vector256.Create(-YCbCrScalar.GCrMult);
-            Vector256<float> bCbMult = Vector256.Create(YCbCrScalar.BCbMult);
-
-            // Walking 8 elements at one step:
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                // y = yVals[i];
-                // cb = cbVals[i] - 128F;
-                // cr = crVals[i] - 128F;
-                ref Vector256<float> c0 = ref Unsafe.Add(ref c0Base, i);
-                ref Vector256<float> c1 = ref Unsafe.Add(ref c1Base, i);
-                ref Vector256<float> c2 = ref Unsafe.Add(ref c2Base, i);
-
-                Vector256<float> y = c0;
-                Vector256<float> cb = c1 + chromaOffset;
-                Vector256<float> cr = c2 + chromaOffset;
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                Vector256<float> r = Vector256_.MultiplyAddEstimate(cr, rCrMult, y);
-                Vector256<float> g = Vector256_.MultiplyAddEstimate(cr, gCrMult, Vector256_.MultiplyAddEstimate(cb, gCbMult, y));
-                Vector256<float> b = Vector256_.MultiplyAddEstimate(cb, bCbMult, y);
-
-                r = Vector256_.RoundToNearestInteger(r) * scale;
-                g = Vector256_.RoundToNearestInteger(g) * scale;
-                b = Vector256_.RoundToNearestInteger(b) * scale;
-
-                c0 = r;
-                c1 = g;
-                c2 = b;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => YCbCrScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector256<float> destY =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> destCb =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> destCr =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            ref Vector256<float> srcR =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector256<float> srcG =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector256<float> srcB =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            Vector256<float> chromaOffset = Vector256.Create(this.HalfValue);
-            Vector256<float> f0299 = Vector256.Create(0.299f);
-            Vector256<float> f0587 = Vector256.Create(0.587f);
-            Vector256<float> f0114 = Vector256.Create(0.114f);
-            Vector256<float> fn0168736 = Vector256.Create(-0.168736f);
-            Vector256<float> fn0331264 = Vector256.Create(-0.331264f);
-            Vector256<float> fn0418688 = Vector256.Create(-0.418688f);
-            Vector256<float> fn0081312F = Vector256.Create(-0.081312F);
-            Vector256<float> f05 = Vector256.Create(0.5f);
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector256<float> r = Unsafe.Add(ref srcR, i);
-                Vector256<float> g = Unsafe.Add(ref srcG, i);
-                Vector256<float> b = Unsafe.Add(ref srcB, i);
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                Vector256<float> y = Vector256_.MultiplyAddEstimate(f0299, r, Vector256_.MultiplyAddEstimate(f0587, g, f0114 * b));
-                Vector256<float> cb = chromaOffset + Vector256_.MultiplyAddEstimate(fn0168736, r, Vector256_.MultiplyAddEstimate(fn0331264, g, f05 * b));
-                Vector256<float> cr = chromaOffset + Vector256_.MultiplyAddEstimate(f05, r, Vector256_.MultiplyAddEstimate(fn0418688, g, fn0081312F * b));
-
-                Unsafe.Add(ref destY, i) = y;
-                Unsafe.Add(ref destCb, i) = cb;
-                Unsafe.Add(ref destCr, i) = cr;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.YCbCrVector512.cs

@@ -1,128 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class YCbCrVector512 : JpegColorConverterVector512
-    {
-        public YCbCrVector512(int precision)
-            : base(JpegColorSpace.YCbCr, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => YCbCrScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceVectorized(in ComponentValues values)
-        {
-            ref Vector512<float> c0Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> c1Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> c2Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            Vector512<float> chromaOffset = Vector512.Create(-this.HalfValue);
-            Vector512<float> scale = Vector512.Create(1 / this.MaximumValue);
-            Vector512<float> rCrMult = Vector512.Create(YCbCrScalar.RCrMult);
-            Vector512<float> gCbMult = Vector512.Create(-YCbCrScalar.GCbMult);
-            Vector512<float> gCrMult = Vector512.Create(-YCbCrScalar.GCrMult);
-            Vector512<float> bCbMult = Vector512.Create(YCbCrScalar.BCbMult);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                // y = yVals[i];
-                // cb = cbVals[i] - 128F;
-                // cr = crVals[i] - 128F;
-                ref Vector512<float> c0 = ref Unsafe.Add(ref c0Base, i);
-                ref Vector512<float> c1 = ref Unsafe.Add(ref c1Base, i);
-                ref Vector512<float> c2 = ref Unsafe.Add(ref c2Base, i);
-
-                Vector512<float> y = c0;
-                Vector512<float> cb = c1 + chromaOffset;
-                Vector512<float> cr = c2 + chromaOffset;
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                Vector512<float> r = Vector512_.MultiplyAddEstimate(cr, rCrMult, y);
-                Vector512<float> g = Vector512_.MultiplyAddEstimate(cr, gCrMult, Vector512_.MultiplyAddEstimate(cb, gCbMult, y));
-                Vector512<float> b = Vector512_.MultiplyAddEstimate(cb, bCbMult, y);
-
-                r = Vector512_.RoundToNearestInteger(r) * scale;
-                g = Vector512_.RoundToNearestInteger(g) * scale;
-                b = Vector512_.RoundToNearestInteger(b) * scale;
-
-                c0 = r;
-                c1 = g;
-                c2 = b;
-            }
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbVectorized(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            ref Vector512<float> destY =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> destCb =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> destCr =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            ref Vector512<float> srcR =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(rLane));
-            ref Vector512<float> srcG =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(gLane));
-            ref Vector512<float> srcB =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(bLane));
-
-            Vector512<float> chromaOffset = Vector512.Create(this.HalfValue);
-            Vector512<float> f0299 = Vector512.Create(0.299f);
-            Vector512<float> f0587 = Vector512.Create(0.587f);
-            Vector512<float> f0114 = Vector512.Create(0.114f);
-            Vector512<float> fn0168736 = Vector512.Create(-0.168736f);
-            Vector512<float> fn0331264 = Vector512.Create(-0.331264f);
-            Vector512<float> fn0418688 = Vector512.Create(-0.418688f);
-            Vector512<float> fn0081312F = Vector512.Create(-0.081312F);
-            Vector512<float> f05 = Vector512.Create(0.5f);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector512<float> r = Unsafe.Add(ref srcR, i);
-                Vector512<float> g = Unsafe.Add(ref srcG, i);
-                Vector512<float> b = Unsafe.Add(ref srcB, i);
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                Vector512<float> y = Vector512_.MultiplyAddEstimate(f0299, r, Vector512_.MultiplyAddEstimate(f0587, g, f0114 * b));
-                Vector512<float> cb = chromaOffset + Vector512_.MultiplyAddEstimate(fn0168736, r, Vector512_.MultiplyAddEstimate(fn0331264, g, f05 * b));
-                Vector512<float> cr = chromaOffset + Vector512_.MultiplyAddEstimate(f05, r, Vector512_.MultiplyAddEstimate(fn0418688, g, fn0081312F * b));
-
-                Unsafe.Add(ref destY, i) = y;
-                Unsafe.Add(ref destCb, i) = cb;
-                Unsafe.Add(ref destCr, i) = cr;
-            }
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceScalarRemainder(in ComponentValues values)
-            => YCbCrScalar.ConvertToRgbInPlace(values, this.MaximumValue, this.HalfValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbScalarRemainder(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => YCbCrScalar.ConvertFromRgb(values, this.HalfValue, rLane, gLane, bLane);
-    }
-}

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

@@ -1,8 +1,13 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
+using System.Numerics;
 using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
+using SixLabors.ImageSharp.ColorProfiles;
+using SixLabors.ImageSharp.ColorProfiles.Icc;
 using SixLabors.ImageSharp.Common.Helpers;
 using SixLabors.ImageSharp.Metadata.Profiles.Icc;
 
@@ -32,9 +37,9 @@ internal abstract partial class JpegColorConverterBase
 
             // YccK first reconstructs inverted RGB in the integer sample domain. Rounding must occur before
             // subtracting from max and applying K because changing that order changes encoded JPEG semantics.
-            c0 = (maximumValue - MathF.Round(y + (YCbCrScalar.RCrMult * cr), MidpointRounding.AwayFromZero)) * scaledK;
+            c0 = (maximumValue - MathF.Round(y + (YCbCrOperator.RCrMult * cr), MidpointRounding.AwayFromZero)) * scaledK;
-            c1 = (maximumValue - MathF.Round(y - (YCbCrScalar.GCbMult * cb) - (YCbCrScalar.GCrMult * cr), MidpointRounding.AwayFromZero)) * scaledK;
+            c1 = (maximumValue - MathF.Round(y - (YCbCrOperator.GCbMult * cb) - (YCbCrOperator.GCrMult * cr), MidpointRounding.AwayFromZero)) * scaledK;
-            c2 = (maximumValue - MathF.Round(y + (YCbCrScalar.BCbMult * cb), MidpointRounding.AwayFromZero)) * scaledK;
+            c2 = (maximumValue - MathF.Round(y + (YCbCrOperator.BCbMult * cb), MidpointRounding.AwayFromZero)) * scaledK;
         }
 
         /// <inheritdoc/>
@@ -47,9 +52,9 @@ internal abstract partial class JpegColorConverterBase
             Vector128<float> scaledK = c3 * scale * scale;
 
             // Four lanes reconstruct YCbCr concurrently; each rounded result is inverted and modulated by its K lane.
-            Vector128<float> r = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(YCbCrScalar.RCrMult), y);
+            Vector128<float> r = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(YCbCrOperator.RCrMult), y);
-            Vector128<float> g = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(-YCbCrScalar.GCrMult), Vector128_.MultiplyAddEstimate(cb, Vector128.Create(-YCbCrScalar.GCbMult), y));
+            Vector128<float> g = Vector128_.MultiplyAddEstimate(cr, Vector128.Create(-YCbCrOperator.GCrMult), Vector128_.MultiplyAddEstimate(cb, Vector128.Create(-YCbCrOperator.GCbMult), y));
-            Vector128<float> b = Vector128_.MultiplyAddEstimate(cb, Vector128.Create(YCbCrScalar.BCbMult), y);
+            Vector128<float> b = Vector128_.MultiplyAddEstimate(cb, Vector128.Create(YCbCrOperator.BCbMult), y);
             c0 = (maximumValue - Vector128_.RoundToNearestInteger(r)) * scaledK;
             c1 = (maximumValue - Vector128_.RoundToNearestInteger(g)) * scaledK;
             c2 = (maximumValue - Vector128_.RoundToNearestInteger(b)) * scaledK;
@@ -65,9 +70,9 @@ internal abstract partial class JpegColorConverterBase
             Vector256<float> scaledK = c3 * scale * scale;
 
             // Eight lanes retain planar alignment from Y/Cb/Cr/K through normalized RGB.
-            Vector256<float> r = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(YCbCrScalar.RCrMult), y);
+            Vector256<float> r = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(YCbCrOperator.RCrMult), y);
-            Vector256<float> g = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(-YCbCrScalar.GCrMult), Vector256_.MultiplyAddEstimate(cb, Vector256.Create(-YCbCrScalar.GCbMult), y));
+            Vector256<float> g = Vector256_.MultiplyAddEstimate(cr, Vector256.Create(-YCbCrOperator.GCrMult), Vector256_.MultiplyAddEstimate(cb, Vector256.Create(-YCbCrOperator.GCbMult), y));
-            Vector256<float> b = Vector256_.MultiplyAddEstimate(cb, Vector256.Create(YCbCrScalar.BCbMult), y);
+            Vector256<float> b = Vector256_.MultiplyAddEstimate(cb, Vector256.Create(YCbCrOperator.BCbMult), y);
             c0 = (maximumValue - Vector256_.RoundToNearestInteger(r)) * scaledK;
             c1 = (maximumValue - Vector256_.RoundToNearestInteger(g)) * scaledK;
             c2 = (maximumValue - Vector256_.RoundToNearestInteger(b)) * scaledK;
@@ -83,9 +88,9 @@ internal abstract partial class JpegColorConverterBase
             Vector512<float> scaledK = c3 * scale * scale;
 
             // Sixteen lanes use the same matrix, rounding, inversion, and K modulation order as scalar code.
-            Vector512<float> r = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(YCbCrScalar.RCrMult), y);
+            Vector512<float> r = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(YCbCrOperator.RCrMult), y);
-            Vector512<float> g = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(-YCbCrScalar.GCrMult), Vector512_.MultiplyAddEstimate(cb, Vector512.Create(-YCbCrScalar.GCbMult), y));
+            Vector512<float> g = Vector512_.MultiplyAddEstimate(cr, Vector512.Create(-YCbCrOperator.GCrMult), Vector512_.MultiplyAddEstimate(cb, Vector512.Create(-YCbCrOperator.GCbMult), y));
-            Vector512<float> b = Vector512_.MultiplyAddEstimate(cb, Vector512.Create(YCbCrScalar.BCbMult), y);
+            Vector512<float> b = Vector512_.MultiplyAddEstimate(cb, Vector512.Create(YCbCrOperator.BCbMult), y);
             c0 = (maximumValue - Vector512_.RoundToNearestInteger(r)) * scaledK;
             c1 = (maximumValue - Vector512_.RoundToNearestInteger(g)) * scaledK;
             c2 = (maximumValue - Vector512_.RoundToNearestInteger(b)) * scaledK;
@@ -130,6 +135,31 @@ internal abstract partial class JpegColorConverterBase
 
         /// <inheritdoc/>
         public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maximumValue)
-            => YccKScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, maximumValue);
+        {
+            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 4);
+            Span<float> packed = memoryOwner.Memory.Span;
+            Span<float> c0 = values.Component0;
+            Span<float> c1 = values.Component1;
+            Span<float> c2 = values.Component2;
+            Span<float> c3 = values.Component3;
+
+            // Adobe-style JPEG YccK is inverted; normalize it before applying the format-defined YccK-to-CMYK transform.
+            PackedInvertNormalizeInterleave4(c0, c1, c2, c3, packed, maximumValue);
+
+            ColorProfileConverter converter = new();
+            Span<Cmyk> source = MemoryMarshal.Cast<float, Cmyk>(packed);
+            converter.Convert<YccK, Cmyk>(MemoryMarshal.Cast<Cmyk, YccK>(source), source);
+
+            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed)[..source.Length];
+            ColorConversionOptions options = new()
+            {
+                SourceIccProfile = profile,
+                TargetIccProfile = CompactSrgbV4Profile.Profile,
+            };
+
+            converter = new ColorProfileConverter(options);
+            converter.Convert<Cmyk, Rgb>(source, destination);
+            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
+        }
     }
 }

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

@@ -1,125 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Buffers;
-using System.Numerics;
-using System.Runtime.InteropServices;
-using SixLabors.ImageSharp.ColorProfiles;
-using SixLabors.ImageSharp.ColorProfiles.Icc;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class YccKScalar : JpegColorConverterScalar
-    {
-        // Derived from ITU-T Rec. T.871
-        internal const float RCrMult = 1.402f;
-        internal const float GCbMult = (float)(0.114 * 1.772 / 0.587);
-        internal const float GCrMult = (float)(0.299 * 1.402 / 0.587);
-        internal const float BCbMult = 1.772f;
-
-        public YccKScalar(int precision)
-            : base(JpegColorSpace.Ycck, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-            => ConvertToRgbInPlace(values, this.MaximumValue, this.HalfValue);
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => ConvertFromRgb(values, this.HalfValue, this.MaximumValue, rLane, gLane, bLane);
-
-        public static void ConvertToRgbInPlace(in ComponentValues values, float maxValue, float halfValue)
-        {
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-            Span<float> c3 = values.Component3;
-
-            float scale = 1 / (maxValue * maxValue);
-
-            for (int i = 0; i < values.Component0.Length; i++)
-            {
-                float y = c0[i];
-                float cb = c1[i] - halfValue;
-                float cr = c2[i] - halfValue;
-                float scaledK = c3[i] * scale;
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                c0[i] = (maxValue - MathF.Round(y + (RCrMult * cr), MidpointRounding.AwayFromZero)) * scaledK;
-                c1[i] = (maxValue - MathF.Round(y - (GCbMult * cb) - (GCrMult * cr), MidpointRounding.AwayFromZero)) * scaledK;
-                c2[i] = (maxValue - MathF.Round(y + (BCbMult * cb), MidpointRounding.AwayFromZero)) * scaledK;
-            }
-        }
-
-        public static void ConvertFromRgb(in ComponentValues values, float halfValue, float maxValue, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            // rgb -> cmyk
-            CmykScalar.ConvertFromRgb(in values, maxValue, rLane, gLane, bLane);
-
-            // cmyk -> ycck
-            Span<float> c = values.Component0;
-            Span<float> m = values.Component1;
-            Span<float> y = values.Component2;
-
-            for (int i = 0; i < y.Length; i++)
-            {
-                float r = maxValue - c[i];
-                float g = maxValue - m[i];
-                float b = maxValue - y[i];
-
-                // k value is passed untouched from rgb -> cmyk conversion
-                c[i] = (0.299f * r) + (0.587f * g) + (0.114f * b);
-                m[i] = halfValue - (0.168736f * r) - (0.331264f * g) + (0.5f * b);
-                y[i] = halfValue + (0.5f * r) - (0.418688f * g) - (0.081312f * b);
-            }
-        }
-
-        public static void ConvertToRgbInPlaceWithIcc(Configuration configuration, IccProfile profile, in ComponentValues values, float maxValue)
-        {
-            using IMemoryOwner<float> memoryOwner = configuration.MemoryAllocator.Allocate<float>(values.Component0.Length * 4);
-            Span<float> packed = memoryOwner.Memory.Span;
-
-            Span<float> c0 = values.Component0;
-            Span<float> c1 = values.Component1;
-            Span<float> c2 = values.Component2;
-            Span<float> c3 = values.Component3;
-
-            PackedInvertNormalizeInterleave4(c0, c1, c2, c3, packed, maxValue);
-
-            ColorProfileConverter converter = new();
-            Span<Cmyk> source = MemoryMarshal.Cast<float, Cmyk>(packed);
-
-            // YccK is not a defined ICC color space — it's a JPEG-specific encoding used in Adobe-style CMYK JPEGs.
-            // ICC profiles expect colorimetric CMYK values, so we must first convert YccK to CMYK using a hardcoded inverse transform.
-            // This transform assumes Rec.601 YCbCr coefficients and an inverted K channel.
-            //
-            // The YccK => Cmyk conversion is independent of any embedded ICC profile.
-            // Since the same RGB working space is used during conversion to and from XYZ,
-            // colorimetric accuracy is preserved.
-            converter.Convert<YccK, Cmyk>(MemoryMarshal.Cast<Cmyk, YccK>(source), source);
-
-            Span<Rgb> destination = MemoryMarshal.Cast<float, Rgb>(packed)[..source.Length];
-
-            ColorConversionOptions options = new()
-            {
-                SourceIccProfile = profile,
-                TargetIccProfile = CompactSrgbV4Profile.Profile,
-            };
-            converter = new ColorProfileConverter(options);
-            converter.Convert<Cmyk, Rgb>(source, destination);
-
-            UnpackDeinterleave3(MemoryMarshal.Cast<float, Vector3>(packed)[..source.Length], c0, c1, c2);
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.YccKVector128.cs

@@ -1,135 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class YccKVector128 : JpegColorConverterVector128
-    {
-        public YccKVector128(int precision)
-            : base(JpegColorSpace.Ycck, 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> kBase =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            // Used for the color conversion
-            Vector128<float> chromaOffset = Vector128.Create(-this.HalfValue);
-            Vector128<float> scale = Vector128.Create(1 / (this.MaximumValue * this.MaximumValue));
-            Vector128<float> max = Vector128.Create(this.MaximumValue);
-            Vector128<float> rCrMult = Vector128.Create(YCbCrScalar.RCrMult);
-            Vector128<float> gCbMult = Vector128.Create(-YCbCrScalar.GCbMult);
-            Vector128<float> gCrMult = Vector128.Create(-YCbCrScalar.GCrMult);
-            Vector128<float> bCbMult = Vector128.Create(YCbCrScalar.BCbMult);
-
-            // Walking 8 elements at one step:
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                // y = yVals[i];
-                // cb = cbVals[i] - 128F;
-                // cr = crVals[i] - 128F;
-                // k = kVals[i] / 256F;
-                ref Vector128<float> c0 = ref Unsafe.Add(ref c0Base, i);
-                ref Vector128<float> c1 = ref Unsafe.Add(ref c1Base, i);
-                ref Vector128<float> c2 = ref Unsafe.Add(ref c2Base, i);
-                Vector128<float> y = c0;
-                Vector128<float> cb = c1 + chromaOffset;
-                Vector128<float> cr = c2 + chromaOffset;
-                Vector128<float> scaledK = Unsafe.Add(ref kBase, i) * scale;
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                Vector128<float> r = Vector128_.MultiplyAddEstimate(cr, rCrMult, y);
-                Vector128<float> g = Vector128_.MultiplyAddEstimate(cr, gCrMult, Vector128_.MultiplyAddEstimate(cb, gCbMult, y));
-                Vector128<float> b = Vector128_.MultiplyAddEstimate(cb, bCbMult, y);
-
-                r = max - Vector128_.RoundToNearestInteger(r);
-                g = max - Vector128_.RoundToNearestInteger(g);
-                b = max - Vector128_.RoundToNearestInteger(b);
-
-                r *= scaledK;
-                g *= scaledK;
-                b *= scaledK;
-
-                c0 = r;
-                c1 = g;
-                c2 = b;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => YccKScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            // rgb -> cmyk
-            CmykVector128.ConvertFromRgb(in values, this.MaximumValue, rLane, gLane, bLane);
-
-            // cmyk -> ycck
-            ref Vector128<float> destY =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector128<float> destCb =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector128<float> destCr =
-                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            ref Vector128<float> srcR = ref destY;
-            ref Vector128<float> srcG = ref destCb;
-            ref Vector128<float> srcB = ref destCr;
-
-            // Used for the color conversion
-            Vector128<float> maxSampleValue = Vector128.Create(this.MaximumValue);
-
-            Vector128<float> chromaOffset = Vector128.Create(this.HalfValue);
-
-            Vector128<float> f0299 = Vector128.Create(0.299f);
-            Vector128<float> f0587 = Vector128.Create(0.587f);
-            Vector128<float> f0114 = Vector128.Create(0.114f);
-            Vector128<float> fn0168736 = Vector128.Create(-0.168736f);
-            Vector128<float> fn0331264 = Vector128.Create(-0.331264f);
-            Vector128<float> fn0418688 = Vector128.Create(-0.418688f);
-            Vector128<float> fn0081312F = Vector128.Create(-0.081312F);
-            Vector128<float> f05 = Vector128.Create(0.5f);
-
-            nuint n = values.Component0.Vector128Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector128<float> r = maxSampleValue - Unsafe.Add(ref srcR, i);
-                Vector128<float> g = maxSampleValue - Unsafe.Add(ref srcG, i);
-                Vector128<float> b = maxSampleValue - Unsafe.Add(ref srcB, i);
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                Vector128<float> y = Vector128_.MultiplyAddEstimate(f0299, r, Vector128_.MultiplyAddEstimate(f0587, g, f0114 * b));
-                Vector128<float> cb = chromaOffset + Vector128_.MultiplyAddEstimate(fn0168736, r, Vector128_.MultiplyAddEstimate(fn0331264, g, f05 * b));
-                Vector128<float> cr = chromaOffset + Vector128_.MultiplyAddEstimate(f05, r, Vector128_.MultiplyAddEstimate(fn0418688, g, fn0081312F * b));
-
-                Unsafe.Add(ref destY, i) = y;
-                Unsafe.Add(ref destCb, i) = cb;
-                Unsafe.Add(ref destCr, i) = cr;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.YccKVector256.cs

@@ -1,135 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class YccKVector256 : JpegColorConverterVector256
-    {
-        public YccKVector256(int precision)
-            : base(JpegColorSpace.Ycck, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlace(in ComponentValues values)
-        {
-            ref Vector256<float> c0Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> c1Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> c2Base =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector256<float> kBase =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            // Used for the color conversion
-            Vector256<float> chromaOffset = Vector256.Create(-this.HalfValue);
-            Vector256<float> scale = Vector256.Create(1 / (this.MaximumValue * this.MaximumValue));
-            Vector256<float> max = Vector256.Create(this.MaximumValue);
-            Vector256<float> rCrMult = Vector256.Create(YCbCrScalar.RCrMult);
-            Vector256<float> gCbMult = Vector256.Create(-YCbCrScalar.GCbMult);
-            Vector256<float> gCrMult = Vector256.Create(-YCbCrScalar.GCrMult);
-            Vector256<float> bCbMult = Vector256.Create(YCbCrScalar.BCbMult);
-
-            // Walking 8 elements at one step:
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                // y = yVals[i];
-                // cb = cbVals[i] - 128F;
-                // cr = crVals[i] - 128F;
-                // k = kVals[i] / 256F;
-                ref Vector256<float> c0 = ref Unsafe.Add(ref c0Base, i);
-                ref Vector256<float> c1 = ref Unsafe.Add(ref c1Base, i);
-                ref Vector256<float> c2 = ref Unsafe.Add(ref c2Base, i);
-                Vector256<float> y = c0;
-                Vector256<float> cb = c1 + chromaOffset;
-                Vector256<float> cr = c2 + chromaOffset;
-                Vector256<float> scaledK = Unsafe.Add(ref kBase, i) * scale;
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                Vector256<float> r = Vector256_.MultiplyAddEstimate(cr, rCrMult, y);
-                Vector256<float> g = Vector256_.MultiplyAddEstimate(cr, gCrMult, Vector256_.MultiplyAddEstimate(cb, gCbMult, y));
-                Vector256<float> b = Vector256_.MultiplyAddEstimate(cb, bCbMult, y);
-
-                r = max - Vector256_.RoundToNearestInteger(r);
-                g = max - Vector256_.RoundToNearestInteger(g);
-                b = max - Vector256_.RoundToNearestInteger(b);
-
-                r *= scaledK;
-                g *= scaledK;
-                b *= scaledK;
-
-                c0 = r;
-                c1 = g;
-                c2 = b;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => YccKScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            // rgb -> cmyk
-            CmykVector256.ConvertFromRgb(in values, this.MaximumValue, rLane, gLane, bLane);
-
-            // cmyk -> ycck
-            ref Vector256<float> destY =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector256<float> destCb =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector256<float> destCr =
-                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            ref Vector256<float> srcR = ref destY;
-            ref Vector256<float> srcG = ref destCb;
-            ref Vector256<float> srcB = ref destCr;
-
-            // Used for the color conversion
-            Vector256<float> maxSampleValue = Vector256.Create(this.MaximumValue);
-
-            Vector256<float> chromaOffset = Vector256.Create(this.HalfValue);
-
-            Vector256<float> f0299 = Vector256.Create(0.299f);
-            Vector256<float> f0587 = Vector256.Create(0.587f);
-            Vector256<float> f0114 = Vector256.Create(0.114f);
-            Vector256<float> fn0168736 = Vector256.Create(-0.168736f);
-            Vector256<float> fn0331264 = Vector256.Create(-0.331264f);
-            Vector256<float> fn0418688 = Vector256.Create(-0.418688f);
-            Vector256<float> fn0081312F = Vector256.Create(-0.081312F);
-            Vector256<float> f05 = Vector256.Create(0.5f);
-
-            nuint n = values.Component0.Vector256Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector256<float> r = maxSampleValue - Unsafe.Add(ref srcR, i);
-                Vector256<float> g = maxSampleValue - Unsafe.Add(ref srcG, i);
-                Vector256<float> b = maxSampleValue - Unsafe.Add(ref srcB, i);
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                Vector256<float> y = Vector256_.MultiplyAddEstimate(f0299, r, Vector256_.MultiplyAddEstimate(f0587, g, f0114 * b));
-                Vector256<float> cb = chromaOffset + Vector256_.MultiplyAddEstimate(fn0168736, r, Vector256_.MultiplyAddEstimate(fn0331264, g, f05 * b));
-                Vector256<float> cr = chromaOffset + Vector256_.MultiplyAddEstimate(f05, r, Vector256_.MultiplyAddEstimate(fn0418688, g, fn0081312F * b));
-
-                Unsafe.Add(ref destY, i) = y;
-                Unsafe.Add(ref destCb, i) = cb;
-                Unsafe.Add(ref destCr, i) = cr;
-            }
-        }
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.YccKVector512.cs

@@ -1,143 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using SixLabors.ImageSharp.Common.Helpers;
-using SixLabors.ImageSharp.Metadata.Profiles.Icc;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    internal sealed class YccKVector512 : JpegColorConverterVector512
-    {
-        public YccKVector512(int precision)
-            : base(JpegColorSpace.Ycck, precision)
-        {
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceVectorized(in ComponentValues values)
-        {
-            ref Vector512<float> c0Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> c1Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> c2Base =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-            ref Vector512<float> kBase =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component3));
-
-            // Used for the color conversion
-            Vector512<float> chromaOffset = Vector512.Create(-this.HalfValue);
-            Vector512<float> scale = Vector512.Create(1 / (this.MaximumValue * this.MaximumValue));
-            Vector512<float> max = Vector512.Create(this.MaximumValue);
-            Vector512<float> rCrMult = Vector512.Create(YCbCrScalar.RCrMult);
-            Vector512<float> gCbMult = Vector512.Create(-YCbCrScalar.GCbMult);
-            Vector512<float> gCrMult = Vector512.Create(-YCbCrScalar.GCrMult);
-            Vector512<float> bCbMult = Vector512.Create(YCbCrScalar.BCbMult);
-
-            // Walking 8 elements at one step:
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                // y = yVals[i];
-                // cb = cbVals[i] - 128F;
-                // cr = crVals[i] - 128F;
-                // k = kVals[i] / 256F;
-                ref Vector512<float> c0 = ref Unsafe.Add(ref c0Base, i);
-                ref Vector512<float> c1 = ref Unsafe.Add(ref c1Base, i);
-                ref Vector512<float> c2 = ref Unsafe.Add(ref c2Base, i);
-                Vector512<float> y = c0;
-                Vector512<float> cb = c1 + chromaOffset;
-                Vector512<float> cr = c2 + chromaOffset;
-                Vector512<float> scaledK = Unsafe.Add(ref kBase, i) * scale;
-
-                // r = y + (1.402F * cr);
-                // g = y - (0.344136F * cb) - (0.714136F * cr);
-                // b = y + (1.772F * cb);
-                Vector512<float> r = Vector512_.MultiplyAddEstimate(cr, rCrMult, y);
-                Vector512<float> g = Vector512_.MultiplyAddEstimate(cr, gCrMult, Vector512_.MultiplyAddEstimate(cb, gCbMult, y));
-                Vector512<float> b = Vector512_.MultiplyAddEstimate(cb, bCbMult, y);
-
-                r = max - Vector512_.RoundToNearestInteger(r);
-                g = max - Vector512_.RoundToNearestInteger(g);
-                b = max - Vector512_.RoundToNearestInteger(b);
-
-                r *= scaledK;
-                g *= scaledK;
-                b *= scaledK;
-
-                c0 = r;
-                c1 = g;
-                c2 = b;
-            }
-        }
-
-        /// <inheritdoc/>
-        public override void ConvertToRgbInPlaceWithIcc(Configuration configuration, in ComponentValues values, IccProfile profile)
-            => YccKScalar.ConvertToRgbInPlaceWithIcc(configuration, profile, values, this.MaximumValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertToRgbInPlaceScalarRemainder(in ComponentValues values)
-            => YccKScalar.ConvertToRgbInPlace(values, this.MaximumValue, this.HalfValue);
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbVectorized(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            // rgb -> cmyk
-            CmykVector512.ConvertFromRgbVectorized(in values, this.MaximumValue, rLane, gLane, bLane);
-
-            // cmyk -> ycck
-            ref Vector512<float> destY =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component0));
-            ref Vector512<float> destCb =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component1));
-            ref Vector512<float> destCr =
-                ref Unsafe.As<float, Vector512<float>>(ref MemoryMarshal.GetReference(values.Component2));
-
-            ref Vector512<float> srcR = ref destY;
-            ref Vector512<float> srcG = ref destCb;
-            ref Vector512<float> srcB = ref destCr;
-
-            // Used for the color conversion
-            Vector512<float> maxSampleValue = Vector512.Create(this.MaximumValue);
-
-            Vector512<float> chromaOffset = Vector512.Create(this.HalfValue);
-
-            Vector512<float> f0299 = Vector512.Create(0.299f);
-            Vector512<float> f0587 = Vector512.Create(0.587f);
-            Vector512<float> f0114 = Vector512.Create(0.114f);
-            Vector512<float> fn0168736 = Vector512.Create(-0.168736f);
-            Vector512<float> fn0331264 = Vector512.Create(-0.331264f);
-            Vector512<float> fn0418688 = Vector512.Create(-0.418688f);
-            Vector512<float> fn0081312F = Vector512.Create(-0.081312F);
-            Vector512<float> f05 = Vector512.Create(0.5f);
-
-            nuint n = values.Component0.Vector512Count<float>();
-            for (nuint i = 0; i < n; i++)
-            {
-                Vector512<float> r = maxSampleValue - Unsafe.Add(ref srcR, i);
-                Vector512<float> g = maxSampleValue - Unsafe.Add(ref srcG, i);
-                Vector512<float> b = maxSampleValue - Unsafe.Add(ref srcB, i);
-
-                // y  =   0 + (0.299 * r) + (0.587 * g) + (0.114 * b)
-                // cb = 128 - (0.168736 * r) - (0.331264 * g) + (0.5 * b)
-                // cr = 128 + (0.5 * r) - (0.418688 * g) - (0.081312 * b)
-                Vector512<float> y = Vector512_.MultiplyAddEstimate(f0299, r, Vector512_.MultiplyAddEstimate(f0587, g, f0114 * b));
-                Vector512<float> cb = chromaOffset + Vector512_.MultiplyAddEstimate(fn0168736, r, Vector512_.MultiplyAddEstimate(fn0331264, g, f05 * b));
-                Vector512<float> cr = chromaOffset + Vector512_.MultiplyAddEstimate(f05, r, Vector512_.MultiplyAddEstimate(fn0418688, g, fn0081312F * b));
-
-                Unsafe.Add(ref destY, i) = y;
-                Unsafe.Add(ref destCb, i) = cb;
-                Unsafe.Add(ref destCr, i) = cr;
-            }
-        }
-
-        /// <inheritdoc/>
-        protected override void ConvertFromRgbScalarRemainder(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-            => YccKScalar.ConvertFromRgb(in values, this.HalfValue, this.MaximumValue, rLane, gLane, bLane);
-    }
-}

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

@@ -1,23 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    /// <summary>
-    /// <see cref="JpegColorConverterBase"/> abstract base for implementations
-    /// based on scalar instructions.
-    /// </summary>
-    internal abstract class JpegColorConverterScalar : JpegColorConverterBase
-    {
-        protected JpegColorConverterScalar(JpegColorSpace colorSpace, int precision)
-            : base(colorSpace, precision)
-        {
-        }
-
-        public sealed override bool IsAvailable => true;
-
-        public sealed override int ElementsPerBatch => 1;
-    }
-}

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

@@ -1,129 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Numerics;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    /// <summary>
-    /// <see cref="JpegColorConverterBase"/> abstract base for implementations
-    /// based on <see cref="Vector"/> API.
-    /// </summary>
-    /// <remarks>
-    /// Converters of this family can work with data of any size.
-    /// Even though real life data is guaranteed to be of size
-    /// divisible by 8 newer SIMD instructions like AVX512 won't work with
-    /// such data out of the box. These converters have fallback code
-    /// for 'remainder' data.
-    /// </remarks>
-    internal abstract class JpegColorConverterVector : JpegColorConverterBase
-    {
-        protected JpegColorConverterVector(JpegColorSpace colorSpace, int precision)
-            : base(colorSpace, precision)
-        {
-        }
-
-        /// <summary>
-        /// Gets a value indicating whether this converter is supported on current hardware.
-        /// </summary>
-        public static bool IsSupported => Vector.IsHardwareAccelerated && Vector<float>.Count % 4 == 0;
-
-        /// <inheritdoc/>
-        public sealed override bool IsAvailable => IsSupported;
-
-        public override int ElementsPerBatch => Vector<float>.Count;
-
-        /// <inheritdoc/>
-        public sealed override void ConvertToRgbInPlace(in ComponentValues values)
-        {
-            DebugGuard.IsTrue(this.IsAvailable, $"{this.GetType().Name} converter is not supported on current hardware.");
-
-            int length = values.Component0.Length;
-            int remainder = (int)((uint)length % (uint)Vector<float>.Count);
-
-            int simdCount = length - remainder;
-            if (simdCount > 0)
-            {
-                this.ConvertToRgbInPlaceVectorized(values.Slice(0, simdCount));
-            }
-
-            // Jpeg images width is always divisible by 8 without a remainder
-            // so it's safe to say SSE/AVX1/AVX2 implementations would never have
-            // 'remainder' pixels
-            // But some exotic simd implementations e.g. AVX-512 can have
-            // remainder pixels
-            if (remainder > 0)
-            {
-                this.ConvertToRgbInPlaceScalarRemainder(values.Slice(simdCount, remainder));
-            }
-        }
-
-        /// <inheritdoc/>
-        public sealed override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            DebugGuard.IsTrue(this.IsAvailable, $"{this.GetType().Name} converter is not supported on current hardware.");
-
-            int length = values.Component0.Length;
-            int remainder = (int)((uint)length % (uint)Vector<float>.Count);
-
-            int simdCount = length - remainder;
-            if (simdCount > 0)
-            {
-                this.ConvertFromRgbVectorized(
-                    values.Slice(0, simdCount),
-                    rLane[..simdCount],
-                    gLane[..simdCount],
-                    bLane[..simdCount]);
-            }
-
-            // Jpeg images width is always divisible by 8 without a remainder
-            // so it's safe to say SSE/AVX1/AVX2 implementations would never have
-            // 'remainder' pixels
-            // But some exotic simd implementations e.g. AVX-512 can have
-            // remainder pixels
-            if (remainder > 0)
-            {
-                this.ConvertFromRgbScalarRemainder(
-                    values.Slice(simdCount, remainder),
-                    rLane.Slice(simdCount, remainder),
-                    gLane.Slice(simdCount, remainder),
-                    bLane.Slice(simdCount, remainder));
-            }
-        }
-
-        /// <summary>
-        /// Converts planar jpeg component values in <paramref name="values"/>
-        /// to RGB color space in place using <see cref="Vector"/> API.
-        /// </summary>
-        /// <param name="values">The input/output as a stack-only <see cref="ComponentValues"/> struct</param>
-        protected abstract void ConvertToRgbInPlaceVectorized(in ComponentValues values);
-
-        /// <summary>
-        /// Converts remainder of the planar jpeg component values after
-        /// conversion in <see cref="ConvertToRgbInPlaceVectorized(in ComponentValues)"/>.
-        /// </summary>
-        /// <param name="values">The input/output as a stack-only <see cref="ComponentValues"/> struct</param>
-        protected abstract void ConvertToRgbInPlaceScalarRemainder(in ComponentValues values);
-
-        /// <summary>
-        /// Converts RGB lanes to jpeg component values using <see cref="Vector"/> API.
-        /// </summary>
-        /// <param name="values">Jpeg component values.</param>
-        /// <param name="rLane">Red colors lane.</param>
-        /// <param name="gLane">Green colors lane.</param>
-        /// <param name="bLane">Blue colors lane.</param>
-        protected abstract void ConvertFromRgbVectorized(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane);
-
-        /// <summary>
-        /// Converts remainder of RGB lanes to jpeg component values after
-        ///  conversion in <see cref="ConvertFromRgbVectorized(in ComponentValues, Span{float}, Span{float}, Span{float})"/>.
-        /// </summary>
-        /// <param name="values">Jpeg component values.</param>
-        /// <param name="rLane">Red colors lane.</param>
-        /// <param name="gLane">Green colors lane.</param>
-        /// <param name="bLane">Blue colors lane.</param>
-        protected abstract void ConvertFromRgbScalarRemainder(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane);
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverterVector128.cs

@@ -1,34 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.Intrinsics;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    /// <summary>
-    /// <see cref="JpegColorConverterBase"/> abstract base for implementations
-    /// based on <see cref="Vector128{T}"/> 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 JpegColorConverterVector128 : JpegColorConverterBase
-    {
-        protected JpegColorConverterVector128(JpegColorSpace colorSpace, int precision)
-            : base(colorSpace, precision)
-        {
-        }
-
-        public static bool IsSupported => Vector128.IsHardwareAccelerated;
-
-        public sealed override bool IsAvailable => IsSupported;
-
-        public sealed override int ElementsPerBatch => Vector128<float>.Count;
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverterVector256.cs

@@ -1,34 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.Intrinsics;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    /// <summary>
-    /// <see cref="JpegColorConverterBase"/> abstract base for implementations
-    /// based on <see cref="Vector256{T}"/> 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 JpegColorConverterVector256 : JpegColorConverterBase
-    {
-        protected JpegColorConverterVector256(JpegColorSpace colorSpace, int precision)
-            : base(colorSpace, precision)
-        {
-        }
-
-        public static bool IsSupported => Vector256.IsHardwareAccelerated;
-
-        public sealed override bool IsAvailable => IsSupported;
-
-        public sealed override int ElementsPerBatch => Vector256<float>.Count;
-    }
-}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverterVector512.cs

@@ -1,111 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Numerics;
-using System.Runtime.Intrinsics;
-
-namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-internal abstract partial class JpegColorConverterBase
-{
-    /// <summary>
-    /// <see cref="JpegColorConverterBase"/> abstract base for implementations
-    /// based on <see cref="Vector512{T}"/> instructions.
-    /// </summary>
-    internal abstract class JpegColorConverterVector512 : JpegColorConverterBase
-    {
-        protected JpegColorConverterVector512(JpegColorSpace colorSpace, int precision)
-            : base(colorSpace, precision)
-        {
-        }
-
-        public static bool IsSupported => Vector512.IsHardwareAccelerated;
-
-        /// <inheritdoc/>
-        public override bool IsAvailable => IsSupported;
-
-        /// <inheritdoc/>
-        public override int ElementsPerBatch => Vector512<float>.Count;
-
-        /// <inheritdoc/>
-        public sealed override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
-        {
-            DebugGuard.IsTrue(this.IsAvailable, $"{this.GetType().Name} converter is not supported on current hardware.");
-
-            int length = values.Component0.Length;
-            int remainder = (int)((uint)length % (uint)Vector512<float>.Count);
-
-            int simdCount = length - remainder;
-            if (simdCount > 0)
-            {
-                this.ConvertFromRgbVectorized(
-                    values.Slice(0, simdCount),
-                    rLane[..simdCount],
-                    gLane[..simdCount],
-                    bLane[..simdCount]);
-            }
-
-            if (remainder > 0)
-            {
-                this.ConvertFromRgbScalarRemainder(
-                    values.Slice(simdCount, remainder),
-                    rLane.Slice(simdCount, remainder),
-                    gLane.Slice(simdCount, remainder),
-                    bLane.Slice(simdCount, remainder));
-            }
-        }
-
-        /// <inheritdoc/>
-        public sealed override void ConvertToRgbInPlace(in ComponentValues values)
-        {
-            DebugGuard.IsTrue(this.IsAvailable, $"{this.GetType().Name} converter is not supported on current hardware.");
-
-            int length = values.Component0.Length;
-            int remainder = (int)((uint)length % (uint)Vector512<float>.Count);
-
-            int simdCount = length - remainder;
-            if (simdCount > 0)
-            {
-                this.ConvertToRgbInPlaceVectorized(values.Slice(0, simdCount));
-            }
-
-            if (remainder > 0)
-            {
-                this.ConvertToRgbInPlaceScalarRemainder(values.Slice(simdCount, remainder));
-            }
-        }
-
-        /// <summary>
-        /// Converts planar jpeg component values in <paramref name="values"/>
-        /// to RGB color space in place using <see cref="Vector"/> API.
-        /// </summary>
-        /// <param name="values">The input/output as a stack-only <see cref="ComponentValues"/> struct</param>
-        protected abstract void ConvertToRgbInPlaceVectorized(in ComponentValues values);
-
-        /// <summary>
-        /// Converts remainder of the planar jpeg component values after
-        /// conversion in <see cref="ConvertToRgbInPlaceVectorized(in ComponentValues)"/>.
-        /// </summary>
-        /// <param name="values">The input/output as a stack-only <see cref="ComponentValues"/> struct</param>
-        protected abstract void ConvertToRgbInPlaceScalarRemainder(in ComponentValues values);
-
-        /// <summary>
-        /// Converts RGB lanes to jpeg component values using <see cref="Vector"/> API.
-        /// </summary>
-        /// <param name="values">Jpeg component values.</param>
-        /// <param name="rLane">Red colors lane.</param>
-        /// <param name="gLane">Green colors lane.</param>
-        /// <param name="bLane">Blue colors lane.</param>
-        protected abstract void ConvertFromRgbVectorized(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane);
-
-        /// <summary>
-        /// Converts remainder of RGB lanes to jpeg component values after
-        ///  conversion in <see cref="ConvertFromRgbVectorized(in ComponentValues, Span{float}, Span{float}, Span{float})"/>.
-        /// </summary>
-        /// <param name="values">Jpeg component values.</param>
-        /// <param name="rLane">Red colors lane.</param>
-        /// <param name="gLane">Green colors lane.</param>
-        /// <param name="bLane">Blue colors lane.</param>
-        protected abstract void ConvertFromRgbScalarRemainder(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane);
-    }
-}

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

@@ -9,40 +9,25 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg;
 [Config(typeof(Config.Short))]
 public class CmykColorConversion : ColorConversionBenchmark
 {
+    private readonly JpegColorConverterBase converter =
+        JpegColorConverterBase.GetConverter(JpegColorSpace.Cmyk, 8);
+
+    /// <summary>
+    /// Initializes a new instance of the <see cref="CmykColorConversion"/> class.
+    /// </summary>
     public CmykColorConversion()
         : base(4)
     {
     }
 
-    [Benchmark(Baseline = true)]
+    /// <summary>
-    public void Scalar()
+    /// Converts one CMYK component row through the adaptive operator traversal.
-    {
+    /// </summary>
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.CmykScalar(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector128()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.CmykVector128(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector256()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.CmykVector256(8).ConvertToRgbInPlace(values);
-    }
-
     [Benchmark]
-    public void SimdVector512()
+    public void ConvertToRgb()
     {
         JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
 
-        new JpegColorConverterBase.CmykVector512(8).ConvertToRgbInPlace(values);
+        this.converter.ConvertToRgbInPlace(values);
     }
 }

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

@@ -9,40 +9,25 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg;
 [Config(typeof(Config.Short))]
 public class GrayScaleColorConversion : ColorConversionBenchmark
 {
+    private readonly JpegColorConverterBase converter =
+        JpegColorConverterBase.GetConverter(JpegColorSpace.Grayscale, 8);
+
+    /// <summary>
+    /// Initializes a new instance of the <see cref="GrayScaleColorConversion"/> class.
+    /// </summary>
     public GrayScaleColorConversion()
         : base(1)
     {
     }
 
-    [Benchmark(Baseline = true)]
+    /// <summary>
-    public void Scalar()
+    /// Converts one grayscale component row through the adaptive operator traversal.
-    {
+    /// </summary>
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.GrayScaleScalar(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector128()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.GrayScaleVector128(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector256()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.GrayScaleVector256(8).ConvertToRgbInPlace(values);
-    }
-
     [Benchmark]
-    public void SimdVector512()
+    public void ConvertToRgb()
     {
         JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
 
-        new JpegColorConverterBase.GrayScaleVector512(8).ConvertToRgbInPlace(values);
+        this.converter.ConvertToRgbInPlace(values);
     }
 }

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

@@ -1,239 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using BenchmarkDotNet.Attributes;
-using BenchmarkDotNet.Columns;
-using BenchmarkDotNet.Configs;
-using SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg;
-
-/// <summary>
-/// Compares each shared operator converter with the Vector512 converter it replaces.
-/// </summary>
-[Config(typeof(Config.Standard))]
-[GroupBenchmarksBy(BenchmarkLogicalGroupRule.ByCategory)]
-[CategoriesColumn]
-public class JpegColorConverterOperatorComparison
-{
-    private JpegColorConverterBase legacy;
-    private JpegColorConverterBase operatorConverter;
-    private float[] legacyC0;
-    private float[] legacyC1;
-    private float[] legacyC2;
-    private float[] legacyC3;
-    private float[] operatorC0;
-    private float[] operatorC1;
-    private float[] operatorC2;
-    private float[] operatorC3;
-    private float[] r;
-    private float[] g;
-    private float[] b;
-    private int componentCount;
-
-    /// <summary>
-    /// Gets or sets the color model measured by the current benchmark case.
-    /// </summary>
-    [Params(
-        JpegColorModel.Grayscale,
-        JpegColorModel.Rgb,
-        JpegColorModel.Cmyk,
-        JpegColorModel.YCbCr,
-        JpegColorModel.YccK,
-        JpegColorModel.TiffCmyk,
-        JpegColorModel.TiffYccK)]
-    public JpegColorModel ColorModel { get; set; }
-
-    /// <summary>
-    /// Gets or sets the number of pixels converted by each invocation.
-    /// </summary>
-    [Params(128, 1024)]
-    public int Count { get; set; }
-
-    /// <summary>
-    /// Creates equivalent legacy and operator converters and their independent component buffers.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        (JpegColorConverterBase Legacy, JpegColorConverterBase Operator, int ComponentCount) converters =
-            this.ColorModel switch
-        {
-            JpegColorModel.Grayscale => (
-                new JpegColorConverterBase.GrayScaleVector512(8),
-                new JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.GrayScaleOperator>(8),
-                1),
-            JpegColorModel.Rgb => (
-                new JpegColorConverterBase.RgbVector512(8),
-                new JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.RgbOperator>(8),
-                3),
-            JpegColorModel.Cmyk => (
-                new JpegColorConverterBase.CmykVector512(8),
-                new JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.CmykOperator>(8),
-                4),
-            JpegColorModel.YCbCr => (
-                new JpegColorConverterBase.YCbCrVector512(8),
-                new JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.YCbCrOperator>(8),
-                3),
-            JpegColorModel.YccK => (
-                new JpegColorConverterBase.YccKVector512(8),
-                new JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.YccKOperator>(8),
-                4),
-            JpegColorModel.TiffCmyk => (
-                new JpegColorConverterBase.TiffCmykVector512(8),
-                new JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.TiffCmykOperator>(8),
-                4),
-            JpegColorModel.TiffYccK => (
-                new JpegColorConverterBase.TiffYccKVector512(8),
-                new JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.TiffYccKOperator>(8),
-                4),
-            _ => throw new InvalidOperationException(),
-        };
-
-        (this.legacy, this.operatorConverter, this.componentCount) = converters;
-
-        Random random = new(42);
-        this.legacyC0 = CreateRandomValues(this.Count, random);
-        this.legacyC1 = CreateRandomValues(this.Count, random);
-        this.legacyC2 = CreateRandomValues(this.Count, random);
-        this.legacyC3 = CreateRandomValues(this.Count, random);
-        this.operatorC0 = this.legacyC0.ToArray();
-        this.operatorC1 = this.legacyC1.ToArray();
-        this.operatorC2 = this.legacyC2.ToArray();
-        this.operatorC3 = this.legacyC3.ToArray();
-        this.r = CreateRandomValues(this.Count, random);
-        this.g = CreateRandomValues(this.Count, random);
-        this.b = CreateRandomValues(this.Count, random);
-    }
-
-    /// <summary>
-    /// Converts JPEG components to RGB using the replaced Vector512 implementation.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("ToRgb")]
-    public void LegacyToRgb()
-    {
-        JpegColorConverterBase.ComponentValues values = this.CreateLegacyValues();
-
-        this.legacy.ConvertToRgbInPlace(values);
-    }
-
-    /// <summary>
-    /// Converts JPEG components to RGB using the shared operator traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("ToRgb")]
-    public void OperatorToRgb()
-    {
-        JpegColorConverterBase.ComponentValues values = this.CreateOperatorValues();
-
-        this.operatorConverter.ConvertToRgbInPlace(values);
-    }
-
-    /// <summary>
-    /// Converts RGB to JPEG components using the replaced Vector512 implementation.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("FromRgb")]
-    public void LegacyFromRgb()
-    {
-        JpegColorConverterBase.ComponentValues values = this.CreateLegacyValues();
-
-        this.legacy.ConvertFromRgb(values, this.r, this.g, this.b);
-    }
-
-    /// <summary>
-    /// Converts RGB to JPEG components using the shared operator traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("FromRgb")]
-    public void OperatorFromRgb()
-    {
-        JpegColorConverterBase.ComponentValues values = this.CreateOperatorValues();
-
-        this.operatorConverter.ConvertFromRgb(values, this.r, this.g, this.b);
-    }
-
-    /// <summary>
-    /// Creates a component view over the buffers owned by the legacy converter.
-    /// </summary>
-    /// <returns>The component view for the configured color model.</returns>
-    private JpegColorConverterBase.ComponentValues CreateLegacyValues()
-        => new(
-            this.componentCount,
-            this.legacyC0,
-            this.componentCount > 1 ? this.legacyC1 : this.legacyC0,
-            this.componentCount > 2 ? this.legacyC2 : this.legacyC0,
-            this.componentCount > 3 ? this.legacyC3 : []);
-
-    /// <summary>
-    /// Creates a component view over the buffers owned by the operator converter.
-    /// </summary>
-    /// <returns>The component view for the configured color model.</returns>
-    private JpegColorConverterBase.ComponentValues CreateOperatorValues()
-        => new(
-            this.componentCount,
-            this.operatorC0,
-            this.componentCount > 1 ? this.operatorC1 : this.operatorC0,
-            this.componentCount > 2 ? this.operatorC2 : this.operatorC0,
-            this.componentCount > 3 ? this.operatorC3 : []);
-
-    /// <summary>
-    /// Creates deterministic sample-domain values for one component plane.
-    /// </summary>
-    /// <param name="length">The number of samples to create.</param>
-    /// <param name="random">The deterministic random source shared by setup.</param>
-    /// <returns>The populated component plane.</returns>
-    private static float[] CreateRandomValues(int length, Random random)
-    {
-        float[] values = new float[length];
-
-        for (int i = 0; i < values.Length; i++)
-        {
-            values[i] = (float)random.NextDouble() * 255F;
-        }
-
-        return values;
-    }
-
-    /// <summary>
-    /// Identifies the JPEG color model used by a benchmark case.
-    /// </summary>
-    public enum JpegColorModel
-    {
-        /// <summary>
-        /// One luminance component.
-        /// </summary>
-        Grayscale,
-
-        /// <summary>
-        /// Three direct RGB components.
-        /// </summary>
-        Rgb,
-
-        /// <summary>
-        /// Four inverted Adobe CMYK components.
-        /// </summary>
-        Cmyk,
-
-        /// <summary>
-        /// Three JPEG YCbCr components.
-        /// </summary>
-        YCbCr,
-
-        /// <summary>
-        /// Four inverted Adobe YCCK components.
-        /// </summary>
-        YccK,
-
-        /// <summary>
-        /// Four non-inverted TIFF CMYK components.
-        /// </summary>
-        TiffCmyk,
-
-        /// <summary>
-        /// Four non-inverted TIFF YCCK components.
-        /// </summary>
-        TiffYccK,
-    }
-}

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

@@ -1,325 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Runtime.CompilerServices;
-using BenchmarkDotNet.Attributes;
-using BenchmarkDotNet.Columns;
-using BenchmarkDotNet.Configs;
-using SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg;
-
-/// <summary>
-/// Exposes every closed JPEG operator traversal beside the Vector512 implementation it replaces.
-/// </summary>
-/// <remarks>
-/// A 63-pixel buffer leaves 256-bit, 128-bit, and scalar remainders after the 512-bit loop, making
-/// every operator overload visible in the generated traversal assembly on AVX-512 hardware.
-/// </remarks>
-[Config(typeof(Config.Analysis))]
-[GroupBenchmarksBy(BenchmarkLogicalGroupRule.ByCategory)]
-[CategoriesColumn]
-public class JpegColorConverterTraversalAssembly
-{
-    private const int Count = 63;
-
-    private readonly JpegColorConverterBase.GrayScaleVector512 grayscaleLegacy = new(8);
-    private readonly JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.GrayScaleOperator> grayscaleOperator = new(8);
-    private readonly JpegColorConverterBase.RgbVector512 rgbLegacy = new(8);
-    private readonly JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.RgbOperator> rgbOperator = new(8);
-    private readonly JpegColorConverterBase.CmykVector512 cmykLegacy = new(8);
-    private readonly JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.CmykOperator> cmykOperator = new(8);
-    private readonly JpegColorConverterBase.YCbCrVector512 yCbCrLegacy = new(8);
-    private readonly JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.YCbCrOperator> yCbCrOperator = new(8);
-    private readonly JpegColorConverterBase.YccKVector512 yccKLegacy = new(8);
-    private readonly JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.YccKOperator> yccKOperator = new(8);
-    private readonly JpegColorConverterBase.TiffCmykVector512 tiffCmykLegacy = new(8);
-    private readonly JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.TiffCmykOperator> tiffCmykOperator = new(8);
-    private readonly JpegColorConverterBase.TiffYccKVector512 tiffYccKLegacy = new(8);
-    private readonly JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.TiffYccKOperator> tiffYccKOperator = new(8);
-
-    private readonly float[] legacyC0 = new float[Count];
-    private readonly float[] legacyC1 = new float[Count];
-    private readonly float[] legacyC2 = new float[Count];
-    private readonly float[] legacyC3 = new float[Count];
-    private readonly float[] operatorC0 = new float[Count];
-    private readonly float[] operatorC1 = new float[Count];
-    private readonly float[] operatorC2 = new float[Count];
-    private readonly float[] operatorC3 = new float[Count];
-    private readonly float[] r = new float[Count];
-    private readonly float[] g = new float[Count];
-    private readonly float[] b = new float[Count];
-
-    /// <summary>
-    /// Populates the component and RGB planes with deterministic sample-domain values.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        Random random = new(42);
-
-        for (int i = 0; i < Count; i++)
-        {
-            // Independent non-constant lanes prevent the JIT from folding arithmetic or mask decisions.
-            this.legacyC0[i] = this.operatorC0[i] = (float)random.NextDouble() * 255F;
-            this.legacyC1[i] = this.operatorC1[i] = (float)random.NextDouble() * 255F;
-            this.legacyC2[i] = this.operatorC2[i] = (float)random.NextDouble() * 255F;
-            this.legacyC3[i] = this.operatorC3[i] = (float)random.NextDouble() * 255F;
-            this.r[i] = (float)random.NextDouble() * 255F;
-            this.g[i] = (float)random.NextDouble() * 255F;
-            this.b[i] = (float)random.NextDouble() * 255F;
-        }
-    }
-
-    /// <summary>
-    /// Runs the replaced grayscale component-to-RGB traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("Grayscale.ToRgb")]
-    public void GrayscaleLegacyToRgb()
-        => this.grayscaleLegacy.ConvertToRgbInPlace(this.CreateLegacyValues(1));
-
-    /// <summary>
-    /// Runs the shared grayscale component-to-RGB traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("Grayscale.ToRgb")]
-    public void GrayscaleOperatorToRgb()
-        => this.grayscaleOperator.ConvertToRgbInPlace(this.CreateOperatorValues(1));
-
-    /// <summary>
-    /// Runs the replaced grayscale RGB-to-component traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("Grayscale.FromRgb")]
-    public void GrayscaleLegacyFromRgb()
-        => this.grayscaleLegacy.ConvertFromRgb(this.CreateLegacyValues(1), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the shared grayscale RGB-to-component traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("Grayscale.FromRgb")]
-    public void GrayscaleOperatorFromRgb()
-        => this.grayscaleOperator.ConvertFromRgb(this.CreateOperatorValues(1), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the replaced RGB component-to-RGB traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("Rgb.ToRgb")]
-    public void RgbLegacyToRgb()
-        => this.rgbLegacy.ConvertToRgbInPlace(this.CreateLegacyValues(3));
-
-    /// <summary>
-    /// Runs the shared RGB component-to-RGB traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("Rgb.ToRgb")]
-    public void RgbOperatorToRgb()
-        => this.rgbOperator.ConvertToRgbInPlace(this.CreateOperatorValues(3));
-
-    /// <summary>
-    /// Runs the replaced RGB RGB-to-component traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("Rgb.FromRgb")]
-    public void RgbLegacyFromRgb()
-        => this.rgbLegacy.ConvertFromRgb(this.CreateLegacyValues(3), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the shared RGB RGB-to-component traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("Rgb.FromRgb")]
-    public void RgbOperatorFromRgb()
-        => this.rgbOperator.ConvertFromRgb(this.CreateOperatorValues(3), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the replaced CMYK component-to-RGB traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("Cmyk.ToRgb")]
-    public void CmykLegacyToRgb()
-        => this.cmykLegacy.ConvertToRgbInPlace(this.CreateLegacyValues(4));
-
-    /// <summary>
-    /// Runs the shared CMYK component-to-RGB traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("Cmyk.ToRgb")]
-    public void CmykOperatorToRgb()
-        => this.cmykOperator.ConvertToRgbInPlace(this.CreateOperatorValues(4));
-
-    /// <summary>
-    /// Runs the replaced CMYK RGB-to-component traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("Cmyk.FromRgb")]
-    public void CmykLegacyFromRgb()
-        => this.cmykLegacy.ConvertFromRgb(this.CreateLegacyValues(4), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the shared CMYK RGB-to-component traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("Cmyk.FromRgb")]
-    public void CmykOperatorFromRgb()
-        => this.cmykOperator.ConvertFromRgb(this.CreateOperatorValues(4), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the replaced YCbCr component-to-RGB traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("YCbCr.ToRgb")]
-    public void YCbCrLegacyToRgb()
-        => this.yCbCrLegacy.ConvertToRgbInPlace(this.CreateLegacyValues(3));
-
-    /// <summary>
-    /// Runs the shared YCbCr component-to-RGB traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("YCbCr.ToRgb")]
-    public void YCbCrOperatorToRgb()
-        => this.yCbCrOperator.ConvertToRgbInPlace(this.CreateOperatorValues(3));
-
-    /// <summary>
-    /// Runs the replaced YCbCr RGB-to-component traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("YCbCr.FromRgb")]
-    public void YCbCrLegacyFromRgb()
-        => this.yCbCrLegacy.ConvertFromRgb(this.CreateLegacyValues(3), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the shared YCbCr RGB-to-component traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("YCbCr.FromRgb")]
-    public void YCbCrOperatorFromRgb()
-        => this.yCbCrOperator.ConvertFromRgb(this.CreateOperatorValues(3), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the replaced YCCK component-to-RGB traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("YccK.ToRgb")]
-    public void YccKLegacyToRgb()
-        => this.yccKLegacy.ConvertToRgbInPlace(this.CreateLegacyValues(4));
-
-    /// <summary>
-    /// Runs the shared YCCK component-to-RGB traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("YccK.ToRgb")]
-    public void YccKOperatorToRgb()
-        => this.yccKOperator.ConvertToRgbInPlace(this.CreateOperatorValues(4));
-
-    /// <summary>
-    /// Runs the replaced YCCK RGB-to-component traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("YccK.FromRgb")]
-    public void YccKLegacyFromRgb()
-        => this.yccKLegacy.ConvertFromRgb(this.CreateLegacyValues(4), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the shared YCCK RGB-to-component traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("YccK.FromRgb")]
-    public void YccKOperatorFromRgb()
-        => this.yccKOperator.ConvertFromRgb(this.CreateOperatorValues(4), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the replaced TIFF CMYK component-to-RGB traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("TiffCmyk.ToRgb")]
-    public void TiffCmykLegacyToRgb()
-        => this.tiffCmykLegacy.ConvertToRgbInPlace(this.CreateLegacyValues(4));
-
-    /// <summary>
-    /// Runs the shared TIFF CMYK component-to-RGB traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("TiffCmyk.ToRgb")]
-    public void TiffCmykOperatorToRgb()
-        => this.tiffCmykOperator.ConvertToRgbInPlace(this.CreateOperatorValues(4));
-
-    /// <summary>
-    /// Runs the replaced TIFF CMYK RGB-to-component traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("TiffCmyk.FromRgb")]
-    public void TiffCmykLegacyFromRgb()
-        => this.tiffCmykLegacy.ConvertFromRgb(this.CreateLegacyValues(4), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the shared TIFF CMYK RGB-to-component traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("TiffCmyk.FromRgb")]
-    public void TiffCmykOperatorFromRgb()
-        => this.tiffCmykOperator.ConvertFromRgb(this.CreateOperatorValues(4), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the replaced TIFF YCCK component-to-RGB traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("TiffYccK.ToRgb")]
-    public void TiffYccKLegacyToRgb()
-        => this.tiffYccKLegacy.ConvertToRgbInPlace(this.CreateLegacyValues(4));
-
-    /// <summary>
-    /// Runs the shared TIFF YCCK component-to-RGB traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("TiffYccK.ToRgb")]
-    public void TiffYccKOperatorToRgb()
-        => this.tiffYccKOperator.ConvertToRgbInPlace(this.CreateOperatorValues(4));
-
-    /// <summary>
-    /// Runs the replaced TIFF YCCK RGB-to-component traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("TiffYccK.FromRgb")]
-    public void TiffYccKLegacyFromRgb()
-        => this.tiffYccKLegacy.ConvertFromRgb(this.CreateLegacyValues(4), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Runs the shared TIFF YCCK RGB-to-component traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("TiffYccK.FromRgb")]
-    public void TiffYccKOperatorFromRgb()
-        => this.tiffYccKOperator.ConvertFromRgb(this.CreateOperatorValues(4), this.r, this.g, this.b);
-
-    /// <summary>
-    /// Creates a correctly aliased component view over the legacy planes.
-    /// </summary>
-    /// <param name="componentCount">The number of component planes owned by the color model.</param>
-    /// <returns>The legacy component view.</returns>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private JpegColorConverterBase.ComponentValues CreateLegacyValues(int componentCount)
-        => new(
-            componentCount,
-            this.legacyC0,
-            componentCount > 1 ? this.legacyC1 : this.legacyC0,
-            componentCount > 2 ? this.legacyC2 : this.legacyC0,
-            componentCount > 3 ? this.legacyC3 : []);
-
-    /// <summary>
-    /// Creates a correctly aliased component view over the operator planes.
-    /// </summary>
-    /// <param name="componentCount">The number of component planes owned by the color model.</param>
-    /// <returns>The operator component view.</returns>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private JpegColorConverterBase.ComponentValues CreateOperatorValues(int componentCount)
-        => new(
-            componentCount,
-            this.operatorC0,
-            componentCount > 1 ? this.operatorC1 : this.operatorC0,
-            componentCount > 2 ? this.operatorC2 : this.operatorC0,
-            componentCount > 3 ? this.operatorC3 : []);
-}

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

@@ -9,40 +9,25 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg;
 [Config(typeof(Config.Short))]
 public class RgbColorConversion : ColorConversionBenchmark
 {
+    private readonly JpegColorConverterBase converter =
+        JpegColorConverterBase.GetConverter(JpegColorSpace.RGB, 8);
+
+    /// <summary>
+    /// Initializes a new instance of the <see cref="RgbColorConversion"/> class.
+    /// </summary>
     public RgbColorConversion()
         : base(3)
     {
     }
 
-    [Benchmark(Baseline = true)]
+    /// <summary>
-    public void Scalar()
+    /// Converts one RGB component row through the adaptive operator traversal.
-    {
+    /// </summary>
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.RgbScalar(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector128()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.RgbVector128(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector256()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.RgbVector256(8).ConvertToRgbInPlace(values);
-    }
-
     [Benchmark]
-    public void SimdVector512()
+    public void ConvertToRgb()
     {
         JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
 
-        new JpegColorConverterBase.RgbVector512(8).ConvertToRgbInPlace(values);
+        this.converter.ConvertToRgbInPlace(values);
     }
 }

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

@@ -9,40 +9,25 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg;
 [Config(typeof(Config.Short))]
 public class YCbCrColorConversion : ColorConversionBenchmark
 {
+    private readonly JpegColorConverterBase converter =
+        JpegColorConverterBase.GetConverter(JpegColorSpace.YCbCr, 8);
+
+    /// <summary>
+    /// Initializes a new instance of the <see cref="YCbCrColorConversion"/> class.
+    /// </summary>
     public YCbCrColorConversion()
         : base(3)
     {
     }
 
+    /// <summary>
+    /// Converts one YCbCr component row through the adaptive operator traversal.
+    /// </summary>
     [Benchmark]
-    public void Scalar()
+    public void ConvertToRgb()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.YCbCrScalar(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector128()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.YCbCrVector128(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector256()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.YCbCrVector256(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector512()
     {
         JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
 
-        new JpegColorConverterBase.YCbCrVector512(8).ConvertToRgbInPlace(values);
+        this.converter.ConvertToRgbInPlace(values);
     }
 }

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

@@ -1,127 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using BenchmarkDotNet.Attributes;
-using BenchmarkDotNet.Columns;
-using BenchmarkDotNet.Configs;
-using SixLabors.ImageSharp.Formats.Jpeg.Components;
-
-namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg;
-
-/// <summary>
-/// Compares the shared YCbCr operator traversal with the Vector512 implementation it replaces.
-/// </summary>
-[GroupBenchmarksBy(BenchmarkLogicalGroupRule.ByCategory)]
-[CategoriesColumn]
-public class YCbCrOperatorComparison
-{
-    private JpegColorConverterBase.YCbCrVector512 legacy;
-    private JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.YCbCrOperator> operatorConverter;
-    private float[] legacyC0;
-    private float[] legacyC1;
-    private float[] legacyC2;
-    private float[] operatorC0;
-    private float[] operatorC1;
-    private float[] operatorC2;
-    private float[] r;
-    private float[] g;
-    private float[] b;
-
-    /// <summary>
-    /// Gets or sets the number of pixels converted by each invocation.
-    /// </summary>
-    [Params(8, 128, 1024)]
-    public int Count { get; set; }
-
-    /// <summary>
-    /// Creates equivalent converter inputs in independent component buffers.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.legacy = new JpegColorConverterBase.YCbCrVector512(8);
-        this.operatorConverter =
-            new JpegColorConverterBase.JpegColorConverter<JpegColorConverterBase.YCbCrOperator>(8);
-
-        Random random = new(42);
-        this.legacyC0 = CreateRandomValues(this.Count, random);
-        this.legacyC1 = CreateRandomValues(this.Count, random);
-        this.legacyC2 = CreateRandomValues(this.Count, random);
-        this.operatorC0 = this.legacyC0.ToArray();
-        this.operatorC1 = this.legacyC1.ToArray();
-        this.operatorC2 = this.legacyC2.ToArray();
-        this.r = CreateRandomValues(this.Count, random);
-        this.g = CreateRandomValues(this.Count, random);
-        this.b = CreateRandomValues(this.Count, random);
-    }
-
-    /// <summary>
-    /// Converts YCbCr components to RGB using the Vector512 implementation.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("ToRgb")]
-    public void LegacyToRgb()
-    {
-        JpegColorConverterBase.ComponentValues values =
-            new(3, this.legacyC0, this.legacyC1, this.legacyC2, []);
-
-        this.legacy.ConvertToRgbInPlace(values);
-    }
-
-    /// <summary>
-    /// Converts YCbCr components to RGB using the shared operator traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("ToRgb")]
-    public void OperatorToRgb()
-    {
-        JpegColorConverterBase.ComponentValues values =
-            new(3, this.operatorC0, this.operatorC1, this.operatorC2, []);
-
-        this.operatorConverter.ConvertToRgbInPlace(values);
-    }
-
-    /// <summary>
-    /// Converts RGB to YCbCr components using the Vector512 implementation.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    [BenchmarkCategory("FromRgb")]
-    public void LegacyFromRgb()
-    {
-        JpegColorConverterBase.ComponentValues values =
-            new(3, this.legacyC0, this.legacyC1, this.legacyC2, []);
-
-        this.legacy.ConvertFromRgb(values, this.r, this.g, this.b);
-    }
-
-    /// <summary>
-    /// Converts RGB to YCbCr components using the shared operator traversal.
-    /// </summary>
-    [Benchmark]
-    [BenchmarkCategory("FromRgb")]
-    public void OperatorFromRgb()
-    {
-        JpegColorConverterBase.ComponentValues values =
-            new(3, this.operatorC0, this.operatorC1, this.operatorC2, []);
-
-        this.operatorConverter.ConvertFromRgb(values, this.r, this.g, this.b);
-    }
-
-    /// <summary>
-    /// Creates deterministic sample-domain values for one component plane.
-    /// </summary>
-    /// <param name="length">The number of samples to create.</param>
-    /// <param name="random">The deterministic random source shared by setup.</param>
-    /// <returns>The populated component plane.</returns>
-    private static float[] CreateRandomValues(int length, Random random)
-    {
-        float[] values = new float[length];
-
-        for (int i = 0; i < values.Length; i++)
-        {
-            values[i] = (float)random.NextDouble() * 255F;
-        }
-
-        return values;
-    }
-}

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

@@ -9,40 +9,25 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg;
 [Config(typeof(Config.Short))]
 public class YccKColorConverter : ColorConversionBenchmark
 {
+    private readonly JpegColorConverterBase converter =
+        JpegColorConverterBase.GetConverter(JpegColorSpace.Ycck, 8);
+
+    /// <summary>
+    /// Initializes a new instance of the <see cref="YccKColorConverter"/> class.
+    /// </summary>
     public YccKColorConverter()
         : base(4)
     {
     }
 
-    [Benchmark(Baseline = true)]
+    /// <summary>
-    public void Scalar()
+    /// Converts one YccK component row through the adaptive operator traversal.
-    {
+    /// </summary>
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.YccKScalar(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector128()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.YccKVector128(8).ConvertToRgbInPlace(values);
-    }
-
-    [Benchmark]
-    public void SimdVector256()
-    {
-        JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
-
-        new JpegColorConverterBase.YccKVector256(8).ConvertToRgbInPlace(values);
-    }
-
     [Benchmark]
-    public void SimdVector512()
+    public void ConvertToRgb()
     {
         JpegColorConverterBase.ComponentValues values = new(this.Input, 0);
 
-        new JpegColorConverterBase.YccKVector512(8).ConvertToRgbInPlace(values);
+        this.converter.ConvertToRgbInPlace(values);
     }
 }

tests/ImageSharp.Benchmarks/Codecs/Png/PngFilterEncode.cs

@@ -8,7 +8,7 @@ using SixLabors.ImageSharp.Formats.Png.Filters;
 namespace SixLabors.ImageSharp.Benchmarks.Codecs.Png;
 
 /// <summary>
-/// Compares the shared PNG map/reduce traversal with the filter-specific traversals it replaces.
+/// Measures the shared PNG filter map/reduce traversal.
 /// </summary>
 [Config(typeof(Config.Short))]
 public class PngFilterEncode
@@ -17,8 +17,7 @@ public class PngFilterEncode
 
     private byte[] scanline;
     private byte[] previousScanline;
-    private byte[] currentResult;
+    private byte[] result;
-    private byte[] baselineResult;
 
     /// <summary>
     /// Gets or sets the filter evaluated by each invocation.
@@ -33,15 +32,14 @@ public class PngFilterEncode
     public int Count { get; set; }
 
     /// <summary>
-    /// Creates deterministic non-uniform inputs and independent result buffers.
+    /// Creates deterministic non-uniform inputs and a result buffer.
     /// </summary>
     [GlobalSetup]
     public void Setup()
     {
         this.scanline = new byte[this.Count];
         this.previousScanline = new byte[this.Count];
-        this.currentResult = new byte[this.Count + 1];
+        this.result = new byte[this.Count + 1];
-        this.baselineResult = new byte[this.Count + 1];
 
         Random random = new(12345678);
         random.NextBytes(this.scanline);
@@ -49,109 +47,53 @@ public class PngFilterEncode
     }
 
     /// <summary>
-    /// Executes the operator-driven map/reduce traversal.
+    /// Executes the shared operator-driven map/reduce traversal.
     /// </summary>
     /// <returns>The filter variance sum.</returns>
     [Benchmark]
-    public int Current()
+    public int Encode()
         => this.Filter switch
         {
-            PngFilterMethod.Sub => this.EncodeSubCurrent(),
+            PngFilterMethod.Sub => this.EncodeSub(),
-            PngFilterMethod.Up => this.EncodeUpCurrent(),
+            PngFilterMethod.Up => this.EncodeUp(),
-            PngFilterMethod.Average => this.EncodeAverageCurrent(),
+            PngFilterMethod.Average => this.EncodeAverage(),
-            PngFilterMethod.Paeth => this.EncodePaethCurrent(),
+            PngFilterMethod.Paeth => this.EncodePaeth(),
             _ => throw new InvalidOperationException()
         };
 
-    /// <summary>
-    /// Executes the filter-specific traversal being replaced.
-    /// </summary>
-    /// <returns>The filter variance sum.</returns>
-    [Benchmark(Baseline = true)]
-    public int Baseline()
-    {
-        int sum;
-
-        switch (this.Filter)
-        {
-            case PngFilterMethod.Sub:
-                PngFilterEncodeBaseline.EncodeSub(
-                    this.scanline,
-                    this.baselineResult,
-                    BytesPerPixel,
-                    out sum);
-
-                break;
-
-            case PngFilterMethod.Up:
-                PngFilterEncodeBaseline.EncodeUp(
-                    this.scanline,
-                    this.previousScanline,
-                    this.baselineResult,
-                    out sum);
-
-                break;
-
-            case PngFilterMethod.Average:
-                PngFilterEncodeBaseline.EncodeAverage(
-                    this.scanline,
-                    this.previousScanline,
-                    this.baselineResult,
-                    BytesPerPixel,
-                    out sum);
-
-                break;
-
-            case PngFilterMethod.Paeth:
-                PngFilterEncodeBaseline.EncodePaeth(
-                    this.scanline,
-                    this.previousScanline,
-                    this.baselineResult,
-                    BytesPerPixel,
-                    out sum);
-
-                break;
-
-            default:
-                throw new InvalidOperationException();
-        }
-
-        return sum;
-    }
-
     /// <summary>
     /// Executes the current Sub encoder.
     /// </summary>
-    private int EncodeSubCurrent()
+    private int EncodeSub()
     {
-        SubFilter.Encode(this.scanline, this.currentResult, BytesPerPixel, out int sum);
+        SubFilter.Encode(this.scanline, this.result, BytesPerPixel, out int sum);
         return sum;
     }
 
     /// <summary>
     /// Executes the current Up encoder.
     /// </summary>
-    private int EncodeUpCurrent()
+    private int EncodeUp()
     {
-        UpFilter.Encode(this.scanline, this.previousScanline, this.currentResult, out int sum);
+        UpFilter.Encode(this.scanline, this.previousScanline, this.result, out int sum);
         return sum;
     }
 
     /// <summary>
     /// Executes the current Average encoder.
     /// </summary>
-    private int EncodeAverageCurrent()
+    private int EncodeAverage()
     {
-        AverageFilter.Encode(this.scanline, this.previousScanline, this.currentResult, BytesPerPixel, out int sum);
+        AverageFilter.Encode(this.scanline, this.previousScanline, this.result, BytesPerPixel, out int sum);
         return sum;
     }
 
     /// <summary>
     /// Executes the current Paeth encoder.
     /// </summary>
-    private int EncodePaethCurrent()
+    private int EncodePaeth()
     {
-        PaethFilter.Encode(this.scanline, this.previousScanline, this.currentResult, BytesPerPixel, out int sum);
+        PaethFilter.Encode(this.scanline, this.previousScanline, this.result, BytesPerPixel, out int sum);
         return sum;
     }
 }

tests/ImageSharp.Benchmarks/Codecs/Png/PngFilterEncodeAssembly.cs

@@ -7,7 +7,7 @@ using SixLabors.ImageSharp.Formats.Png.Filters;
 namespace SixLabors.ImageSharp.Benchmarks.Codecs.Png;
 
 /// <summary>
-/// Exposes every normalized PNG filter and retained baseline for assembly comparison.
+/// Exposes every normalized PNG filter for assembly inspection.
 /// </summary>
 [Config(typeof(Config.Analysis))]
 public class PngFilterEncodeAssembly
@@ -17,8 +17,7 @@ public class PngFilterEncodeAssembly
 
     private byte[] scanline;
     private byte[] previousScanline;
-    private byte[] currentResult;
+    private byte[] result;
-    private byte[] baselineResult;
 
     /// <summary>
     /// Creates inputs whose suffix exercises 512-, 256-, and 128-bit register widths.
@@ -28,8 +27,7 @@ public class PngFilterEncodeAssembly
     {
         this.scanline = new byte[Count];
         this.previousScanline = new byte[Count];
-        this.currentResult = new byte[Count + 1];
+        this.result = new byte[Count + 1];
-        this.baselineResult = new byte[Count + 1];
 
         Random random = new(12345678);
         random.NextBytes(this.scanline);
@@ -41,64 +39,26 @@ public class PngFilterEncodeAssembly
     /// </summary>
     [Benchmark]
     public void Sub()
-        => SubFilter.Encode(this.scanline, this.currentResult, BytesPerPixel, out _);
+        => SubFilter.Encode(this.scanline, this.result, BytesPerPixel, out _);
 
     /// <summary>
     /// Executes the normalized Up encoder.
     /// </summary>
     [Benchmark]
     public void Up()
-        => UpFilter.Encode(this.scanline, this.previousScanline, this.currentResult, out _);
+        => UpFilter.Encode(this.scanline, this.previousScanline, this.result, out _);
 
     /// <summary>
     /// Executes the normalized Average encoder.
     /// </summary>
     [Benchmark]
     public void Average()
-        => AverageFilter.Encode(this.scanline, this.previousScanline, this.currentResult, BytesPerPixel, out _);
+        => AverageFilter.Encode(this.scanline, this.previousScanline, this.result, BytesPerPixel, out _);
 
     /// <summary>
     /// Executes the normalized Paeth encoder.
     /// </summary>
     [Benchmark]
     public void Paeth()
-        => PaethFilter.Encode(this.scanline, this.previousScanline, this.currentResult, BytesPerPixel, out _);
+        => PaethFilter.Encode(this.scanline, this.previousScanline, this.result, BytesPerPixel, out _);
-
-    /// <summary>
-    /// Executes the retained Sub encoder.
-    /// </summary>
-    [Benchmark]
-    public void BaselineSub()
-        => PngFilterEncodeBaseline.EncodeSub(this.scanline, this.baselineResult, BytesPerPixel, out _);
-
-    /// <summary>
-    /// Executes the retained Up encoder.
-    /// </summary>
-    [Benchmark]
-    public void BaselineUp()
-        => PngFilterEncodeBaseline.EncodeUp(this.scanline, this.previousScanline, this.baselineResult, out _);
-
-    /// <summary>
-    /// Executes the retained Average encoder.
-    /// </summary>
-    [Benchmark]
-    public void BaselineAverage()
-        => PngFilterEncodeBaseline.EncodeAverage(
-            this.scanline,
-            this.previousScanline,
-            this.baselineResult,
-            BytesPerPixel,
-            out _);
-
-    /// <summary>
-    /// Executes the retained Paeth encoder.
-    /// </summary>
-    [Benchmark]
-    public void BaselinePaeth()
-        => PngFilterEncodeBaseline.EncodePaeth(
-            this.scanline,
-            this.previousScanline,
-            this.baselineResult,
-            BytesPerPixel,
-            out _);
 }

tests/ImageSharp.Benchmarks/Codecs/Png/PngFilterEncodeBaseline.cs

@@ -1,470 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Numerics;
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using System.Runtime.Intrinsics.X86;
-
-namespace SixLabors.ImageSharp.Benchmarks.Codecs.Png;
-
-/// <summary>
-/// Retains the filter-specific PNG encode traversals for direct performance comparison.
-/// </summary>
-internal static class PngFilterEncodeBaseline
-{
-    /// <summary>
-    /// Executes the filter-specific Sub traversal.
-    /// </summary>
-    public static void EncodeSub(
-        ReadOnlySpan<byte> scanline,
-        Span<byte> result,
-        int bytesPerPixel,
-        out int sum)
-    {
-        ref byte scanBaseRef = ref MemoryMarshal.GetReference(scanline);
-        ref byte resultBaseRef = ref MemoryMarshal.GetReference(result);
-        sum = 0;
-        resultBaseRef = 1;
-
-        nuint x = 0;
-
-        for (; x < (uint)bytesPerPixel;)
-        {
-            byte scan = Unsafe.Add(ref scanBaseRef, x);
-            x++;
-            ref byte residual = ref Unsafe.Add(ref resultBaseRef, x);
-            residual = scan;
-            sum += Numerics.Abs(unchecked((sbyte)residual));
-        }
-
-        if (Avx2.IsSupported)
-        {
-            Vector256<byte> zero = Vector256<byte>.Zero;
-            Vector256<int> accumulator = Vector256<int>.Zero;
-
-            for (nuint xLeft = x - (uint)bytesPerPixel; (int)x <= scanline.Length - Vector256<byte>.Count; xLeft += (uint)Vector256<byte>.Count)
-            {
-                Vector256<byte> scan = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref scanBaseRef, x));
-                Vector256<byte> left = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref scanBaseRef, xLeft));
-                Vector256<byte> residual = Avx2.Subtract(scan, left);
-
-                Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref resultBaseRef, x + 1)) = residual;
-                x += (uint)Vector256<byte>.Count;
-                accumulator = Avx2.Add(
-                    accumulator,
-                    Avx2.SumAbsoluteDifferences(Avx2.Abs(residual.AsSByte()), zero).AsInt32());
-            }
-
-            sum += Numerics.EvenReduceSum(accumulator);
-        }
-        else if (Vector.IsHardwareAccelerated)
-        {
-            Vector<uint> accumulator = Vector<uint>.Zero;
-
-            for (nuint xLeft = x - (uint)bytesPerPixel; (int)x <= scanline.Length - Vector<byte>.Count; xLeft += (uint)Vector<byte>.Count)
-            {
-                Vector<byte> scan = Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref scanBaseRef, x));
-                Vector<byte> left = Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref scanBaseRef, xLeft));
-                Vector<byte> residual = scan - left;
-
-                Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref resultBaseRef, x + 1)) = residual;
-                x += (uint)Vector<byte>.Count;
-                Numerics.Accumulate(
-                    ref accumulator,
-                    Vector.AsVectorByte(Vector.Abs(Vector.AsVectorSByte(residual))));
-            }
-
-            for (int i = 0; i < Vector<uint>.Count; i++)
-            {
-                sum += (int)accumulator[i];
-            }
-        }
-
-        for (nuint xLeft = x - (uint)bytesPerPixel; x < (uint)scanline.Length; xLeft++)
-        {
-            byte scan = Unsafe.Add(ref scanBaseRef, x);
-            byte left = Unsafe.Add(ref scanBaseRef, xLeft);
-            x++;
-            ref byte residual = ref Unsafe.Add(ref resultBaseRef, x);
-            residual = (byte)(scan - left);
-            sum += Numerics.Abs(unchecked((sbyte)residual));
-        }
-    }
-
-    /// <summary>
-    /// Executes the filter-specific Up traversal.
-    /// </summary>
-    public static void EncodeUp(
-        ReadOnlySpan<byte> scanline,
-        ReadOnlySpan<byte> previousScanline,
-        Span<byte> result,
-        out int sum)
-    {
-        ref byte scanBaseRef = ref MemoryMarshal.GetReference(scanline);
-        ref byte previousBaseRef = ref MemoryMarshal.GetReference(previousScanline);
-        ref byte resultBaseRef = ref MemoryMarshal.GetReference(result);
-        sum = 0;
-        resultBaseRef = 2;
-
-        nuint x = 0;
-
-        if (Avx2.IsSupported)
-        {
-            Vector256<byte> zero = Vector256<byte>.Zero;
-            Vector256<int> accumulator = Vector256<int>.Zero;
-
-            for (; (int)x <= scanline.Length - Vector256<byte>.Count;)
-            {
-                Vector256<byte> scan = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref scanBaseRef, x));
-                Vector256<byte> above = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref previousBaseRef, x));
-                Vector256<byte> residual = Avx2.Subtract(scan, above);
-
-                Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref resultBaseRef, x + 1)) = residual;
-                x += (uint)Vector256<byte>.Count;
-                accumulator = Avx2.Add(
-                    accumulator,
-                    Avx2.SumAbsoluteDifferences(Avx2.Abs(residual.AsSByte()), zero).AsInt32());
-            }
-
-            sum += Numerics.EvenReduceSum(accumulator);
-        }
-        else if (Vector.IsHardwareAccelerated)
-        {
-            Vector<uint> accumulator = Vector<uint>.Zero;
-
-            for (; (int)x <= scanline.Length - Vector<byte>.Count;)
-            {
-                Vector<byte> scan = Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref scanBaseRef, x));
-                Vector<byte> above = Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref previousBaseRef, x));
-                Vector<byte> residual = scan - above;
-
-                Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref resultBaseRef, x + 1)) = residual;
-                x += (uint)Vector<byte>.Count;
-                Numerics.Accumulate(
-                    ref accumulator,
-                    Vector.AsVectorByte(Vector.Abs(Vector.AsVectorSByte(residual))));
-            }
-
-            for (int i = 0; i < Vector<uint>.Count; i++)
-            {
-                sum += (int)accumulator[i];
-            }
-        }
-
-        for (; x < (uint)scanline.Length;)
-        {
-            byte scan = Unsafe.Add(ref scanBaseRef, x);
-            byte above = Unsafe.Add(ref previousBaseRef, x);
-            x++;
-            ref byte residual = ref Unsafe.Add(ref resultBaseRef, x);
-            residual = (byte)(scan - above);
-            sum += Numerics.Abs(unchecked((sbyte)residual));
-        }
-    }
-
-    /// <summary>
-    /// Executes the filter-specific Average traversal.
-    /// </summary>
-    public static void EncodeAverage(
-        ReadOnlySpan<byte> scanline,
-        ReadOnlySpan<byte> previousScanline,
-        Span<byte> result,
-        uint bytesPerPixel,
-        out int sum)
-    {
-        ref byte scanBaseRef = ref MemoryMarshal.GetReference(scanline);
-        ref byte previousBaseRef = ref MemoryMarshal.GetReference(previousScanline);
-        ref byte resultBaseRef = ref MemoryMarshal.GetReference(result);
-        sum = 0;
-        resultBaseRef = 3;
-
-        nuint x = 0;
-
-        for (; x < bytesPerPixel;)
-        {
-            byte scan = Unsafe.Add(ref scanBaseRef, x);
-            byte above = Unsafe.Add(ref previousBaseRef, x);
-            x++;
-            ref byte residual = ref Unsafe.Add(ref resultBaseRef, x);
-            residual = (byte)(scan - (above >> 1));
-            sum += Numerics.Abs(unchecked((sbyte)residual));
-        }
-
-        if (Avx2.IsSupported)
-        {
-            Vector256<byte> zero = Vector256<byte>.Zero;
-            Vector256<int> accumulator = Vector256<int>.Zero;
-            Vector256<byte> allBitsSet = Avx2.CompareEqual(accumulator, accumulator).AsByte();
-
-            for (nuint xLeft = x - bytesPerPixel; (int)x <= scanline.Length - Vector256<byte>.Count; xLeft += (uint)Vector256<byte>.Count)
-            {
-                Vector256<byte> scan = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref scanBaseRef, x));
-                Vector256<byte> left = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref scanBaseRef, xLeft));
-                Vector256<byte> above = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref previousBaseRef, x));
-                Vector256<byte> average = Avx2.Xor(
-                    Avx2.Average(Avx2.Xor(left, allBitsSet), Avx2.Xor(above, allBitsSet)),
-                    allBitsSet);
-
-                Vector256<byte> residual = Avx2.Subtract(scan, average);
-                Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref resultBaseRef, x + 1)) = residual;
-                x += (uint)Vector256<byte>.Count;
-                accumulator = Avx2.Add(
-                    accumulator,
-                    Avx2.SumAbsoluteDifferences(Avx2.Abs(residual.AsSByte()), zero).AsInt32());
-            }
-
-            sum += Numerics.EvenReduceSum(accumulator);
-        }
-        else if (Sse2.IsSupported)
-        {
-            Vector128<byte> zero = Vector128<byte>.Zero;
-            Vector128<int> accumulator = Vector128<int>.Zero;
-            Vector128<byte> allBitsSet = Sse2.CompareEqual(accumulator, accumulator).AsByte();
-
-            for (nuint xLeft = x - bytesPerPixel; (int)x <= scanline.Length - Vector128<byte>.Count; xLeft += (uint)Vector128<byte>.Count)
-            {
-                Vector128<byte> scan = Unsafe.As<byte, Vector128<byte>>(ref Unsafe.Add(ref scanBaseRef, x));
-                Vector128<byte> left = Unsafe.As<byte, Vector128<byte>>(ref Unsafe.Add(ref scanBaseRef, xLeft));
-                Vector128<byte> above = Unsafe.As<byte, Vector128<byte>>(ref Unsafe.Add(ref previousBaseRef, x));
-                Vector128<byte> average = Sse2.Xor(
-                    Sse2.Average(Sse2.Xor(left, allBitsSet), Sse2.Xor(above, allBitsSet)),
-                    allBitsSet);
-
-                Vector128<byte> residual = Sse2.Subtract(scan, average);
-                Unsafe.As<byte, Vector128<byte>>(ref Unsafe.Add(ref resultBaseRef, x + 1)) = residual;
-                x += (uint)Vector128<byte>.Count;
-
-                Vector128<byte> absolute;
-
-                if (Ssse3.IsSupported)
-                {
-                    absolute = Ssse3.Abs(residual.AsSByte());
-                }
-                else
-                {
-                    Vector128<sbyte> mask = Sse2.CompareGreaterThan(zero.AsSByte(), residual.AsSByte());
-                    absolute = Sse2.Xor(Sse2.Add(residual.AsSByte(), mask), mask).AsByte();
-                }
-
-                accumulator = Sse2.Add(
-                    accumulator,
-                    Sse2.SumAbsoluteDifferences(absolute, zero).AsInt32());
-            }
-
-            sum += Numerics.EvenReduceSum(accumulator);
-        }
-
-        for (nuint xLeft = x - bytesPerPixel; x < (uint)scanline.Length; xLeft++)
-        {
-            byte scan = Unsafe.Add(ref scanBaseRef, x);
-            byte left = Unsafe.Add(ref scanBaseRef, xLeft);
-            byte above = Unsafe.Add(ref previousBaseRef, x);
-            x++;
-            ref byte residual = ref Unsafe.Add(ref resultBaseRef, x);
-            residual = (byte)(scan - ((left + above) >> 1));
-            sum += Numerics.Abs(unchecked((sbyte)residual));
-        }
-    }
-
-    /// <summary>
-    /// Executes the filter-specific Paeth traversal.
-    /// </summary>
-    public static void EncodePaeth(
-        ReadOnlySpan<byte> scanline,
-        ReadOnlySpan<byte> previousScanline,
-        Span<byte> result,
-        int bytesPerPixel,
-        out int sum)
-    {
-        ref byte scanBaseRef = ref MemoryMarshal.GetReference(scanline);
-        ref byte previousBaseRef = ref MemoryMarshal.GetReference(previousScanline);
-        ref byte resultBaseRef = ref MemoryMarshal.GetReference(result);
-        sum = 0;
-        resultBaseRef = 4;
-
-        nuint x = 0;
-
-        for (; x < (uint)bytesPerPixel;)
-        {
-            byte scan = Unsafe.Add(ref scanBaseRef, x);
-            byte above = Unsafe.Add(ref previousBaseRef, x);
-            x++;
-            ref byte residual = ref Unsafe.Add(ref resultBaseRef, x);
-            residual = (byte)(scan - above);
-            sum += Numerics.Abs(unchecked((sbyte)residual));
-        }
-
-        if (Avx2.IsSupported)
-        {
-            Vector256<byte> zero = Vector256<byte>.Zero;
-            Vector256<int> accumulator = Vector256<int>.Zero;
-
-            for (nuint xLeft = x - (uint)bytesPerPixel; (int)x <= scanline.Length - Vector256<byte>.Count; xLeft += (uint)Vector256<byte>.Count)
-            {
-                Vector256<byte> scan = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref scanBaseRef, x));
-                Vector256<byte> left = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref scanBaseRef, xLeft));
-                Vector256<byte> above = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref previousBaseRef, x));
-                Vector256<byte> upperLeft = Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref previousBaseRef, xLeft));
-                Vector256<byte> residual = Avx2.Subtract(scan, PaethPredictor(left, above, upperLeft));
-
-                Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref resultBaseRef, x + 1)) = residual;
-                x += (uint)Vector256<byte>.Count;
-                accumulator = Avx2.Add(
-                    accumulator,
-                    Avx2.SumAbsoluteDifferences(Avx2.Abs(residual.AsSByte()), zero).AsInt32());
-            }
-
-            sum += Numerics.EvenReduceSum(accumulator);
-        }
-        else if (Vector.IsHardwareAccelerated)
-        {
-            Vector<uint> accumulator = Vector<uint>.Zero;
-
-            for (nuint xLeft = x - (uint)bytesPerPixel; (int)x <= scanline.Length - Vector<byte>.Count; xLeft += (uint)Vector<byte>.Count)
-            {
-                Vector<byte> scan = Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref scanBaseRef, x));
-                Vector<byte> left = Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref scanBaseRef, xLeft));
-                Vector<byte> above = Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref previousBaseRef, x));
-                Vector<byte> upperLeft = Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref previousBaseRef, xLeft));
-                Vector<byte> residual = scan - PaethPredictor(left, above, upperLeft);
-
-                Unsafe.As<byte, Vector<byte>>(ref Unsafe.Add(ref resultBaseRef, x + 1)) = residual;
-                x += (uint)Vector<byte>.Count;
-                Numerics.Accumulate(
-                    ref accumulator,
-                    Vector.AsVectorByte(Vector.Abs(Vector.AsVectorSByte(residual))));
-            }
-
-            for (int i = 0; i < Vector<uint>.Count; i++)
-            {
-                sum += (int)accumulator[i];
-            }
-        }
-
-        for (nuint xLeft = x - (uint)bytesPerPixel; x < (uint)scanline.Length; xLeft++)
-        {
-            byte scan = Unsafe.Add(ref scanBaseRef, x);
-            byte left = Unsafe.Add(ref scanBaseRef, xLeft);
-            byte above = Unsafe.Add(ref previousBaseRef, x);
-            byte upperLeft = Unsafe.Add(ref previousBaseRef, xLeft);
-            x++;
-            ref byte residual = ref Unsafe.Add(ref resultBaseRef, x);
-            residual = (byte)(scan - PaethPredictor(left, above, upperLeft));
-            sum += Numerics.Abs(unchecked((sbyte)residual));
-        }
-    }
-
-    /// <summary>
-    /// Selects the scalar Paeth predictor.
-    /// </summary>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static byte PaethPredictor(byte left, byte above, byte upperLeft)
-    {
-        int p = left + above - upperLeft;
-        int distanceLeft = Numerics.Abs(p - left);
-        int distanceAbove = Numerics.Abs(p - above);
-        int distanceUpperLeft = Numerics.Abs(p - upperLeft);
-
-        if (distanceLeft <= distanceAbove && distanceLeft <= distanceUpperLeft)
-        {
-            return left;
-        }
-
-        return distanceAbove <= distanceUpperLeft ? above : upperLeft;
-    }
-
-    /// <summary>
-    /// Selects the AVX2 Paeth predictor.
-    /// </summary>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static Vector256<byte> PaethPredictor(
-        Vector256<byte> left,
-        Vector256<byte> above,
-        Vector256<byte> upperLeft)
-    {
-        Vector256<byte> zero = Vector256<byte>.Zero;
-        Vector256<byte> aboveMinusUpper = Avx2.SubtractSaturate(above, upperLeft);
-        Vector256<byte> leftMinusUpper = Avx2.SubtractSaturate(left, upperLeft);
-        Vector256<byte> distanceLeft =
-            Avx2.Or(Avx2.SubtractSaturate(upperLeft, above), aboveMinusUpper);
-
-        Vector256<byte> distanceAbove =
-            Avx2.Or(Avx2.SubtractSaturate(upperLeft, left), leftMinusUpper);
-
-        Vector256<byte> sameDirection = Avx2.CompareEqual(
-            Avx2.CompareEqual(aboveMinusUpper, zero),
-            Avx2.CompareEqual(leftMinusUpper, zero));
-
-        Vector256<byte> distanceUpper = Avx2.Or(
-            sameDirection,
-            Avx2.Or(
-                Avx2.SubtractSaturate(distanceAbove, distanceLeft),
-                Avx2.SubtractSaturate(distanceLeft, distanceAbove)));
-
-        Vector256<byte> minimumAboveUpper = Avx2.Min(distanceUpper, distanceAbove);
-        Vector256<byte> aboveOrUpper = Avx2.BlendVariable(
-            upperLeft,
-            above,
-            Avx2.CompareEqual(minimumAboveUpper, distanceAbove));
-
-        return Avx2.BlendVariable(
-            aboveOrUpper,
-            left,
-            Avx2.CompareEqual(Avx2.Min(minimumAboveUpper, distanceLeft), distanceLeft));
-    }
-
-    /// <summary>
-    /// Selects the portable vector Paeth predictor.
-    /// </summary>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static Vector<byte> PaethPredictor(
-        Vector<byte> left,
-        Vector<byte> above,
-        Vector<byte> upperLeft)
-    {
-        Vector.Widen(left, out Vector<ushort> leftLow, out Vector<ushort> leftHigh);
-        Vector.Widen(above, out Vector<ushort> aboveLow, out Vector<ushort> aboveHigh);
-        Vector.Widen(upperLeft, out Vector<ushort> upperLow, out Vector<ushort> upperHigh);
-
-        Vector<short> lower = PaethPredictor(
-            Vector.AsVectorInt16(leftLow),
-            Vector.AsVectorInt16(aboveLow),
-            Vector.AsVectorInt16(upperLow));
-
-        Vector<short> upper = PaethPredictor(
-            Vector.AsVectorInt16(leftHigh),
-            Vector.AsVectorInt16(aboveHigh),
-            Vector.AsVectorInt16(upperHigh));
-
-        return Vector.AsVectorByte(Vector.Narrow(lower, upper));
-    }
-
-    /// <summary>
-    /// Selects the portable widened Paeth predictor.
-    /// </summary>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static Vector<short> PaethPredictor(
-        Vector<short> left,
-        Vector<short> above,
-        Vector<short> upperLeft)
-    {
-        Vector<short> p = left + above - upperLeft;
-        Vector<short> distanceLeft = Vector.Abs(p - left);
-        Vector<short> distanceAbove = Vector.Abs(p - above);
-        Vector<short> distanceUpper = Vector.Abs(p - upperLeft);
-
-        Vector<short> chooseLeft = Vector.BitwiseAnd(
-            Vector.LessThanOrEqual(distanceLeft, distanceAbove),
-            Vector.LessThanOrEqual(distanceLeft, distanceUpper));
-
-        return Vector.ConditionalSelect(
-            chooseLeft,
-            left,
-            Vector.ConditionalSelect(
-                Vector.LessThanOrEqual(distanceAbove, distanceUpper),
-                above,
-                upperLeft));
-    }
-}

tests/ImageSharp.Benchmarks/General/BasicMath/AddSpan.cs

@@ -1,65 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using BenchmarkDotNet.Attributes;
-using SixLabors.ImageSharp.Common.Helpers;
-
-namespace SixLabors.ImageSharp.Benchmarks.General.BasicMath;
-
-public class AddSpan
-{
-    private byte[] scalarValues = null!;
-    private byte[] tensorValues = null!;
-    private byte[] addends = null!;
-
-    /// <summary>
-    /// Gets or sets the number of values to add.
-    /// </summary>
-    [Params(32, 257, 2048)]
-    public int Length { get; set; }
-
-    /// <summary>
-    /// Creates equivalent deterministic inputs for both implementations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.scalarValues = new byte[this.Length];
-        this.tensorValues = new byte[this.Length];
-        this.addends = new byte[this.Length];
-
-        for (int i = 0; i < this.Length; i++)
-        {
-            byte value = (byte)((i * 17) + 31);
-            this.scalarValues[i] = value;
-            this.tensorValues[i] = value;
-            this.addends[i] = (byte)((i * 29) + 7);
-        }
-    }
-
-    /// <summary>
-    /// Adds the values with a scalar loop.
-    /// </summary>
-    /// <returns>The first result, which keeps the mutated data observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public byte Scalar()
-    {
-        for (int i = 0; i < this.scalarValues.Length; i++)
-        {
-            this.scalarValues[i] += this.addends[i];
-        }
-
-        return this.scalarValues[0];
-    }
-
-    /// <summary>
-    /// Adds the values with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the mutated data observable to the benchmark harness.</returns>
-    [Benchmark]
-    public byte TensorPipeline()
-    {
-        TensorPrimitives_.Add<byte>(this.tensorValues, this.addends, this.tensorValues);
-        return this.tensorValues[0];
-    }
-}

tests/ImageSharp.Benchmarks/General/BasicMath/NormalizeSpan.cs

@@ -1,61 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using BenchmarkDotNet.Attributes;
-
-namespace SixLabors.ImageSharp.Benchmarks.General.BasicMath;
-
-public class NormalizeSpan
-{
-    private float[] scalarValues = null!;
-    private float[] tensorValues = null!;
-
-    /// <summary>
-    /// Gets or sets the number of values to normalize.
-    /// </summary>
-    [Params(7, 32, 257, 2048)]
-    public int Length { get; set; }
-
-    /// <summary>
-    /// Creates equivalent deterministic inputs for both implementations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.scalarValues = new float[this.Length];
-        this.tensorValues = new float[this.Length];
-
-        for (int i = 0; i < this.scalarValues.Length; i++)
-        {
-            float value = ((i * 17) % 251) + 1;
-            this.scalarValues[i] = value;
-            this.tensorValues[i] = value;
-        }
-    }
-
-    /// <summary>
-    /// Normalizes the values with a scalar loop.
-    /// </summary>
-    /// <returns>The first result, which keeps the mutated data observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public float Scalar()
-    {
-        for (int i = 0; i < this.scalarValues.Length; i++)
-        {
-            this.scalarValues[i] /= 4096F;
-        }
-
-        return this.scalarValues[0];
-    }
-
-    /// <summary>
-    /// Normalizes the values with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the mutated data observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float TensorPipeline()
-    {
-        Numerics.Normalize(this.tensorValues, 4096F);
-        return this.tensorValues[0];
-    }
-}

tests/ImageSharp.Benchmarks/General/BasicMath/TensorPrimitivesAssembly.cs

@@ -0,0 +1,175 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+using System.Numerics;
+using BenchmarkDotNet.Attributes;
+using SixLabors.ImageSharp.Common.Helpers;
+
+namespace SixLabors.ImageSharp.Benchmarks.General.BasicMath;
+
+/// <summary>
+/// Exposes every floating-point tensor compatibility operation for assembly inspection.
+/// </summary>
+[Config(typeof(Config.Analysis))]
+public class TensorPrimitivesAssembly
+{
+    private const int Count = 2048;
+
+    private readonly float[] x = new float[Count];
+    private readonly float[] y = new float[Count];
+    private readonly float[] destination = new float[Count];
+
+    /// <summary>
+    /// Populates the input spans with deterministic non-uniform values.
+    /// </summary>
+    [GlobalSetup]
+    public void Setup()
+    {
+        for (int i = 0; i < Count; i++)
+        {
+            this.x[i] = ((i * 17) % 251) + 1;
+            this.y[i] = ((i * 29) % 251) + 1;
+        }
+    }
+
+    /// <summary>
+    /// Adds two floating-point spans.
+    /// </summary>
+    /// <returns>The first result, which keeps the destination observable.</returns>
+    [Benchmark]
+    public float Add()
+    {
+        TensorPrimitives_.Add<float>(this.x, this.y, this.destination);
+        return this.destination[0];
+    }
+
+    /// <summary>
+    /// Clamps a floating-point span between scalar bounds.
+    /// </summary>
+    /// <returns>The first result, which keeps the destination observable.</returns>
+    [Benchmark]
+    public float Clamp()
+    {
+        TensorPrimitives_.Clamp(this.x, 64F, 128F, this.destination);
+        return this.destination[0];
+    }
+
+    /// <summary>
+    /// Divides a floating-point span by a scalar.
+    /// </summary>
+    /// <returns>The first result, which keeps the destination observable.</returns>
+    [Benchmark]
+    public float Divide()
+    {
+        TensorPrimitives_.Divide(this.x, 4096F, this.destination);
+        return this.destination[0];
+    }
+
+    /// <summary>
+    /// Computes the element-wise maximum of a floating-point span and a scalar.
+    /// </summary>
+    /// <returns>The first result, which keeps the destination observable.</returns>
+    [Benchmark]
+    public float Max()
+    {
+        TensorPrimitives_.Max(this.x, 64F, this.destination);
+        return this.destination[0];
+    }
+
+    /// <summary>
+    /// Multiplies a floating-point span by a scalar.
+    /// </summary>
+    /// <returns>The first result, which keeps the destination observable.</returns>
+    [Benchmark]
+    public float Multiply()
+    {
+        TensorPrimitives_.Multiply(this.x, 0.5F, this.destination);
+        return this.destination[0];
+    }
+}
+
+/// <summary>
+/// Exposes integral addition specializations for assembly inspection.
+/// </summary>
+/// <typeparam name="T">The integral element type.</typeparam>
+[Config(typeof(Config.Analysis))]
+[GenericTypeArguments(typeof(byte))]
+[GenericTypeArguments(typeof(uint))]
+public class TensorPrimitivesIntegralAddAssembly<T>
+    where T : unmanaged, INumber<T>
+{
+    private const int Count = 2048;
+
+    private readonly T[] x = new T[Count];
+    private readonly T[] y = new T[Count];
+    private readonly T[] destination = new T[Count];
+
+    /// <summary>
+    /// Populates the input spans with deterministic non-uniform values.
+    /// </summary>
+    [GlobalSetup]
+    public void Setup()
+    {
+        for (int i = 0; i < Count; i++)
+        {
+            this.x[i] = T.CreateTruncating((i * 17) + 31);
+            this.y[i] = T.CreateTruncating((i * 29) + 7);
+        }
+    }
+
+    /// <summary>
+    /// Adds two integral spans.
+    /// </summary>
+    /// <returns>The first result, which keeps the destination observable.</returns>
+    [Benchmark]
+    public T Add()
+    {
+        TensorPrimitives_.Add<T>(this.x, this.y, this.destination);
+        return this.destination[0];
+    }
+}
+
+/// <summary>
+/// Exposes integral clamp specializations for assembly inspection.
+/// </summary>
+/// <typeparam name="T">The integral element type.</typeparam>
+[Config(typeof(Config.Analysis))]
+[GenericTypeArguments(typeof(byte))]
+[GenericTypeArguments(typeof(uint))]
+[GenericTypeArguments(typeof(int))]
+public class TensorPrimitivesIntegralClampAssembly<T>
+    where T : unmanaged, INumber<T>
+{
+    private const int Count = 2048;
+
+    private readonly T[] source = new T[Count];
+    private readonly T[] destination = new T[Count];
+    private T min;
+    private T max;
+
+    /// <summary>
+    /// Populates the input span and scalar bounds with deterministic values.
+    /// </summary>
+    [GlobalSetup]
+    public void Setup()
+    {
+        this.min = T.CreateTruncating(64);
+        this.max = T.CreateTruncating(128);
+
+        for (int i = 0; i < Count; i++)
+        {
+            this.source[i] = T.CreateTruncating((i * 31) % 257);
+        }
+    }
+
+    /// <summary>
+    /// Clamps an integral span between scalar bounds.
+    /// </summary>
+    /// <returns>The first result, which keeps the destination observable.</returns>
+    [Benchmark]
+    public T Clamp()
+    {
+        TensorPrimitives_.Clamp(this.source, this.min, this.max, this.destination);
+        return this.destination[0];
+    }
+}

tests/ImageSharp.Benchmarks/General/BasicMath/TensorPrimitivesAssemblyComparison.cs

@@ -1,813 +0,0 @@
-// Copyright (c) Six Labors.
-// Licensed under the Six Labors Split License.
-
-using System.Numerics;
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
-using System.Runtime.Intrinsics.X86;
-using BenchmarkDotNet.Attributes;
-using SixLabors.ImageSharp.Common.Helpers;
-
-namespace SixLabors.ImageSharp.Benchmarks.General.BasicMath;
-
-#pragma warning disable SA1649 // File name should match first type name
-public class TensorPrimitivesJpegMultiplyAssemblyComparison
-#pragma warning restore SA1649 // File name should match first type name
-{
-    private readonly float multiplier = -1F;
-    private float[] legacyValues = null!;
-    private float[] tensorValues = null!;
-
-    /// <summary>
-    /// Creates equivalent stable inputs for both implementations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.legacyValues = new float[256];
-        this.tensorValues = new float[256];
-
-        for (int i = 0; i < this.legacyValues.Length; i++)
-        {
-            float value = ((i * 17) % 251) + 1;
-            this.legacyValues[i] = value;
-            this.tensorValues[i] = value;
-        }
-    }
-
-    /// <summary>
-    /// Multiplies the row with the retired JPEG AVX pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public float Legacy()
-    {
-        LegacyMultiply(this.legacyValues, this.multiplier);
-        return this.legacyValues[0];
-    }
-
-    /// <summary>
-    /// Multiplies the row with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float Tensor()
-    {
-        TensorPrimitives_.Multiply(this.tensorValues, this.multiplier, this.tensorValues);
-        return this.tensorValues[0];
-    }
-
-    /// <summary>
-    /// Reproduces the retired JPEG multiplication loop for assembly comparison.
-    /// </summary>
-    /// <param name="target">The row to multiply.</param>
-    /// <param name="multiplier">The scalar multiplier.</param>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static void LegacyMultiply(Span<float> target, float multiplier)
-    {
-        ref Vector256<float> targetVector = ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(target));
-        nuint count = (uint)target.Length / (uint)Vector256<float>.Count;
-        Vector256<float> multiplierVector = Vector256.Create(multiplier);
-
-        for (nuint i = 0; i < count; i++)
-        {
-            Unsafe.Add(ref targetVector, i) = Avx.Multiply(Unsafe.Add(ref targetVector, i), multiplierVector);
-        }
-    }
-}
-
-public class TensorPrimitivesNormalizeAssemblyComparison
-{
-    private readonly float divisor = -1F;
-    private float[] legacyValues = null!;
-    private float[] tensorValues = null!;
-
-    /// <summary>
-    /// Creates equivalent stable inputs for both implementations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.legacyValues = new float[7];
-        this.tensorValues = new float[7];
-
-        for (int i = 0; i < this.legacyValues.Length; i++)
-        {
-            float value = ((i * 17) % 251) + 1;
-            this.legacyValues[i] = value;
-            this.tensorValues[i] = value;
-        }
-    }
-
-    /// <summary>
-    /// Normalizes the values with the retired fixed-width pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public float Legacy()
-    {
-        LegacyNormalize(this.legacyValues, this.divisor);
-        return this.legacyValues[0];
-    }
-
-    /// <summary>
-    /// Normalizes the values with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float Tensor()
-    {
-        Numerics.Normalize(this.tensorValues, this.divisor);
-        return this.tensorValues[0];
-    }
-
-    /// <summary>
-    /// Reproduces the retired normalization loop for assembly comparison.
-    /// </summary>
-    /// <param name="span">The values to normalize.</param>
-    /// <param name="sum">The scalar divisor.</param>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static void LegacyNormalize(Span<float> span, float sum)
-    {
-        ref float start = ref MemoryMarshal.GetReference(span);
-        ref float vectorEnd = ref Unsafe.Add(ref start, span.Length & ~7);
-        Vector256<float> sum256 = Vector256.Create(sum);
-
-        while (Unsafe.IsAddressLessThan(ref start, ref vectorEnd))
-        {
-            Unsafe.As<float, Vector256<float>>(ref start) /= sum256;
-            start = ref Unsafe.Add(ref start, (nuint)8);
-        }
-
-        if ((span.Length & 7) >= 4)
-        {
-            Unsafe.As<float, Vector128<float>>(ref start) /= sum256.GetLower();
-            start = ref Unsafe.Add(ref start, (nuint)4);
-        }
-
-        ref float end = ref Unsafe.Add(ref start, span.Length & 3);
-
-        while (Unsafe.IsAddressLessThan(ref start, ref end))
-        {
-            start /= sum;
-            start = ref Unsafe.Add(ref start, (nuint)1);
-        }
-    }
-}
-
-public class TensorPrimitivesUInt32AssemblyComparison
-{
-    private uint[] x = null!;
-    private uint[] y = null!;
-    private uint[] legacyDestination = null!;
-    private uint[] tensorDestination = null!;
-
-    /// <summary>
-    /// Creates deterministic histogram inputs and independent destinations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.x = new uint[2048];
-        this.y = new uint[2048];
-        this.legacyDestination = new uint[2048];
-        this.tensorDestination = new uint[2048];
-
-        for (int i = 0; i < this.x.Length; i++)
-        {
-            this.x[i] = (uint)((i * 17) + 31);
-            this.y[i] = (uint)((i * 29) + 7);
-        }
-    }
-
-    /// <summary>
-    /// Adds histogram bins with the retired four-vector AVX2 pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public uint Legacy()
-    {
-        LegacyAdd(this.x, this.y, this.legacyDestination);
-        return this.legacyDestination[0];
-    }
-
-    /// <summary>
-    /// Adds histogram bins with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public uint Tensor()
-    {
-        TensorPrimitives_.Add<uint>(this.x, this.y, this.tensorDestination);
-        return this.tensorDestination[0];
-    }
-
-    /// <summary>
-    /// Reproduces the retired WebP histogram addition loop for assembly comparison.
-    /// </summary>
-    /// <param name="x">The first histogram.</param>
-    /// <param name="y">The second histogram.</param>
-    /// <param name="destination">The destination receiving the sums.</param>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static void LegacyAdd(ReadOnlySpan<uint> x, ReadOnlySpan<uint> y, Span<uint> destination)
-    {
-        ref uint xRef = ref MemoryMarshal.GetReference(x);
-        ref uint yRef = ref MemoryMarshal.GetReference(y);
-        ref uint destinationRef = ref MemoryMarshal.GetReference(destination);
-
-        nuint index = 0;
-
-        do
-        {
-            Vector256<uint> x0 = Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref xRef, index));
-            Vector256<uint> x1 = Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref xRef, index + 8));
-            Vector256<uint> x2 = Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref xRef, index + 16));
-            Vector256<uint> x3 = Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref xRef, index + 24));
-            Vector256<uint> y0 = Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref yRef, index));
-            Vector256<uint> y1 = Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref yRef, index + 8));
-            Vector256<uint> y2 = Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref yRef, index + 16));
-            Vector256<uint> y3 = Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref yRef, index + 24));
-
-            Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref destinationRef, index)) = Avx2.Add(x0, y0);
-            Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref destinationRef, index + 8)) = Avx2.Add(x1, y1);
-            Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref destinationRef, index + 16)) = Avx2.Add(x2, y2);
-            Unsafe.As<uint, Vector256<uint>>(ref Unsafe.Add(ref destinationRef, index + 24)) = Avx2.Add(x3, y3);
-            index += 32;
-        }
-        while (index <= (uint)x.Length - 32);
-
-        for (int i = (int)index; i < x.Length; i++)
-        {
-            destination[i] = x[i] + y[i];
-        }
-    }
-}
-
-public class TensorPrimitivesByteAssemblyComparison
-{
-    private byte[] x = null!;
-    private byte[] y = null!;
-    private byte[] legacyDestination = null!;
-    private byte[] tensorDestination = null!;
-
-    /// <summary>
-    /// Creates deterministic byte inputs and independent destinations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.x = new byte[2048];
-        this.y = new byte[2048];
-        this.legacyDestination = new byte[2048];
-        this.tensorDestination = new byte[2048];
-
-        for (int i = 0; i < this.x.Length; i++)
-        {
-            this.x[i] = (byte)((i * 17) + 31);
-            this.y[i] = (byte)((i * 29) + 7);
-        }
-    }
-
-    /// <summary>
-    /// Adds bytes with the retired WebP AVX2 pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public byte Legacy()
-    {
-        LegacyAdd(this.x, this.y, this.legacyDestination);
-        return this.legacyDestination[0];
-    }
-
-    /// <summary>
-    /// Adds bytes with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public byte Tensor()
-    {
-        TensorPrimitives_.Add<byte>(this.x, this.y, this.tensorDestination);
-        return this.tensorDestination[0];
-    }
-
-    /// <summary>
-    /// Reproduces the retired WebP byte addition loop for assembly comparison.
-    /// </summary>
-    /// <param name="x">The first input.</param>
-    /// <param name="y">The second input.</param>
-    /// <param name="destination">The destination receiving modulo-256 sums.</param>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static void LegacyAdd(ReadOnlySpan<byte> x, ReadOnlySpan<byte> y, Span<byte> destination)
-    {
-        ref byte xRef = ref MemoryMarshal.GetReference(x);
-        ref byte yRef = ref MemoryMarshal.GetReference(y);
-        ref byte destinationRef = ref MemoryMarshal.GetReference(destination);
-
-        nuint i;
-        int maxPosition = x.Length & ~31;
-
-        for (i = 0; i < (uint)maxPosition; i += 32)
-        {
-            Vector256<int> x0 = Unsafe.As<byte, Vector256<int>>(ref Unsafe.Add(ref xRef, i));
-            Vector256<int> y0 = Unsafe.As<byte, Vector256<int>>(ref Unsafe.Add(ref yRef, i));
-            Vector256<byte> result = x0.AsByte() + y0.AsByte();
-            Unsafe.As<byte, Vector256<byte>>(ref Unsafe.Add(ref destinationRef, i)) = result;
-        }
-
-        for (; i < (uint)x.Length; i++)
-        {
-            Unsafe.Add(ref destinationRef, i) = (byte)(Unsafe.Add(ref xRef, i) + Unsafe.Add(ref yRef, i));
-        }
-    }
-}
-
-public class TensorPrimitivesSingleAddAssemblyComparison
-{
-    private float[] legacyTarget = null!;
-    private float[] tensorTarget = null!;
-    private float[] source = null!;
-
-    /// <summary>
-    /// Creates deterministic JPEG row inputs.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.legacyTarget = new float[2048];
-        this.tensorTarget = new float[2048];
-        this.source = new float[2048];
-
-        for (int i = 0; i < this.source.Length; i++)
-        {
-            float value = ((i * 17) % 251) + 1;
-            this.legacyTarget[i] = value;
-            this.tensorTarget[i] = value;
-            this.source[i] = ((i * 29) % 31) - 15;
-        }
-    }
-
-    /// <summary>
-    /// Adds JPEG row values with the retired AVX pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public float Legacy()
-    {
-        LegacyAdd(this.legacyTarget, this.source);
-        return this.legacyTarget[0];
-    }
-
-    /// <summary>
-    /// Adds JPEG row values with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float Tensor()
-    {
-        TensorPrimitives_.Add<float>(this.tensorTarget, this.source, this.tensorTarget);
-        return this.tensorTarget[0];
-    }
-
-    /// <summary>
-    /// Reproduces the retired JPEG row addition loop for assembly comparison.
-    /// </summary>
-    /// <param name="target">The destination row.</param>
-    /// <param name="source">The row added to the destination.</param>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static void LegacyAdd(Span<float> target, ReadOnlySpan<float> source)
-    {
-        ref Vector256<float> targetVector = ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(target));
-        ref Vector256<float> sourceVector = ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(source));
-        nuint count = (uint)source.Length / (uint)Vector256<float>.Count;
-
-        for (nuint i = 0; i < count; i++)
-        {
-            Unsafe.Add(ref targetVector, i) = Avx.Add(Unsafe.Add(ref targetVector, i), Unsafe.Add(ref sourceVector, i));
-        }
-    }
-}
-
-[GenericTypeArguments(typeof(byte))]
-[GenericTypeArguments(typeof(uint))]
-[GenericTypeArguments(typeof(int))]
-[GenericTypeArguments(typeof(float))]
-[GenericTypeArguments(typeof(double))]
-public class TensorPrimitivesClampAssemblyComparison<T>
-    where T : unmanaged, INumber<T>
-{
-    private T[] legacyValues = null!;
-    private T[] tensorValues = null!;
-    private T min;
-    private T max;
-
-    /// <summary>
-    /// Creates deterministic clamp inputs for the current element type.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.legacyValues = new T[2048];
-        this.tensorValues = new T[2048];
-        this.min = T.CreateTruncating(64);
-        this.max = T.CreateTruncating(128);
-
-        for (int i = 0; i < this.legacyValues.Length; i++)
-        {
-            T value = T.CreateTruncating((i * 31) % 257);
-            this.legacyValues[i] = value;
-            this.tensorValues[i] = value;
-        }
-    }
-
-    /// <summary>
-    /// Clamps values with the retired <see cref="Vector{T}"/> pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public T Legacy()
-    {
-        LegacyClamp(this.legacyValues, this.min, this.max);
-        return this.legacyValues[0];
-    }
-
-    /// <summary>
-    /// Clamps values with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public T Tensor()
-    {
-        TensorPrimitives_.Clamp(this.tensorValues, this.min, this.max, this.tensorValues);
-        return this.tensorValues[0];
-    }
-
-    /// <summary>
-    /// Reproduces the retired clamp pipeline for assembly comparison.
-    /// </summary>
-    /// <param name="span">The values to clamp.</param>
-    /// <param name="min">The inclusive lower bound.</param>
-    /// <param name="max">The inclusive upper bound.</param>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static void LegacyClamp(Span<T> span, T min, T max)
-    {
-        int remainder = Numerics.ModuloP2(span.Length, Vector<T>.Count);
-        int adjustedCount = span.Length - remainder;
-
-        if (adjustedCount > 0)
-        {
-            Vector<T> vectorMin = new(min);
-            Vector<T> vectorMax = new(max);
-            nint vectorCount = (nint)(uint)adjustedCount / Vector<T>.Count;
-            nint remainingVectors = Numerics.Modulo4(vectorCount);
-            nint unrolledVectors = vectorCount - remainingVectors;
-
-            ref Vector<T> current0 = ref Unsafe.As<T, Vector<T>>(ref MemoryMarshal.GetReference(span));
-            ref Vector<T> current1 = ref Unsafe.Add(ref current0, 1);
-            ref Vector<T> current2 = ref Unsafe.Add(ref current0, 2);
-            ref Vector<T> current3 = ref Unsafe.Add(ref current0, 3);
-            ref Vector<T> end = ref Unsafe.Add(ref current0, unrolledVectors);
-
-            while (Unsafe.IsAddressLessThan(ref current0, ref end))
-            {
-                current0 = Vector.Min(Vector.Max(vectorMin, current0), vectorMax);
-                current1 = Vector.Min(Vector.Max(vectorMin, current1), vectorMax);
-                current2 = Vector.Min(Vector.Max(vectorMin, current2), vectorMax);
-                current3 = Vector.Min(Vector.Max(vectorMin, current3), vectorMax);
-
-                current0 = ref Unsafe.Add(ref current0, 4);
-                current1 = ref Unsafe.Add(ref current1, 4);
-                current2 = ref Unsafe.Add(ref current2, 4);
-                current3 = ref Unsafe.Add(ref current3, 4);
-            }
-
-            if (remainingVectors > 0)
-            {
-                current0 = ref end;
-                end = ref Unsafe.Add(ref end, remainingVectors);
-
-                while (Unsafe.IsAddressLessThan(ref current0, ref end))
-                {
-                    current0 = Vector.Min(Vector.Max(vectorMin, current0), vectorMax);
-                    current0 = ref Unsafe.Add(ref current0, 1);
-                }
-            }
-        }
-
-        for (int i = adjustedCount; i < span.Length; i++)
-        {
-            T value = span[i];
-            span[i] = value > max ? max : value < min ? min : value;
-        }
-    }
-}
-
-public class TensorPrimitivesIccMaxAssemblyComparison
-{
-    private Vector4[] legacyValues = null!;
-    private Vector4[] tensorValues = null!;
-
-    /// <summary>
-    /// Creates deterministic ICC values containing positive and negative channels.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.legacyValues = new Vector4[512];
-        this.tensorValues = new Vector4[512];
-
-        for (int i = 0; i < this.legacyValues.Length; i++)
-        {
-            float value = ((i * 17) % 251) - 125;
-            Vector4 vector = new(value, value + 1, value - 1, value + 2);
-            this.legacyValues[i] = vector;
-            this.tensorValues[i] = vector;
-        }
-    }
-
-    /// <summary>
-    /// Clips negative channels with the retired ICC pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public float Legacy()
-    {
-        for (int i = 0; i < this.legacyValues.Length; i++)
-        {
-            this.legacyValues[i] = Vector4.Max(this.legacyValues[i], Vector4.Zero);
-        }
-
-        return this.legacyValues[0].X;
-    }
-
-    /// <summary>
-    /// Clips negative channels with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float Tensor()
-    {
-        Span<float> values = MemoryMarshal.Cast<Vector4, float>(this.tensorValues.AsSpan());
-        TensorPrimitives_.Max(values, 0F, values);
-        return values[0];
-    }
-}
-
-public class TensorPrimitivesIccMultiplyAssemblyComparison
-{
-    private readonly float multiplier = 65280F / 65535F;
-    private Vector4[] source = null!;
-    private Vector4[] legacyDestination = null!;
-    private Vector4[] tensorDestination = null!;
-
-    /// <summary>
-    /// Creates deterministic ICC inputs and independent destinations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.source = new Vector4[512];
-        this.legacyDestination = new Vector4[512];
-        this.tensorDestination = new Vector4[512];
-
-        for (int i = 0; i < this.source.Length; i++)
-        {
-            float value = ((i * 17) % 251) + 1;
-            this.source[i] = new Vector4(value, value + 1, value + 2, value + 3);
-        }
-    }
-
-    /// <summary>
-    /// Multiplies ICC channels with the retired <see cref="Vector{T}"/> pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public float Legacy()
-    {
-        Span<float> source = MemoryMarshal.Cast<Vector4, float>(this.source.AsSpan());
-        Span<float> destination = MemoryMarshal.Cast<Vector4, float>(this.legacyDestination.AsSpan());
-        ref Vector<float> sourceVector = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(source));
-        ref Vector<float> destinationVector = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(destination));
-        Vector<float> scale = new(this.multiplier);
-        nuint count = (uint)source.Length / (uint)Vector<float>.Count;
-
-        for (nuint i = 0; i < count; i++)
-        {
-            Unsafe.Add(ref destinationVector, i) = Unsafe.Add(ref sourceVector, i) * scale;
-        }
-
-        return destination[0];
-    }
-
-    /// <summary>
-    /// Multiplies ICC channels with the tensor compatibility pipeline.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float Tensor()
-    {
-        Span<float> source = MemoryMarshal.Cast<Vector4, float>(this.source.AsSpan());
-        Span<float> destination = MemoryMarshal.Cast<Vector4, float>(this.tensorDestination.AsSpan());
-        TensorPrimitives_.Multiply(source, this.multiplier, destination);
-        return destination[0];
-    }
-}
-
-#if NET10_0_OR_GREATER
-[GenericTypeArguments(typeof(byte))]
-[GenericTypeArguments(typeof(uint))]
-[GenericTypeArguments(typeof(float))]
-public class TensorPrimitivesRuntimeAddAssemblyComparison<T>
-    where T : unmanaged, INumber<T>
-{
-    private T[] x = null!;
-    private T[] y = null!;
-    private T[] compatibilityDestination = null!;
-    private T[] runtimeDestination = null!;
-
-    /// <summary>
-    /// Creates deterministic inputs and independent destinations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.x = new T[2048];
-        this.y = new T[2048];
-        this.compatibilityDestination = new T[2048];
-        this.runtimeDestination = new T[2048];
-
-        for (int i = 0; i < this.x.Length; i++)
-        {
-            this.x[i] = T.CreateTruncating((i * 17) + 31);
-            this.y[i] = T.CreateTruncating((i * 29) + 7);
-        }
-    }
-
-    /// <summary>
-    /// Adds values with the compatibility implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public T Compatibility()
-    {
-        TensorPrimitives_.Add<T>(this.x, this.y, this.compatibilityDestination);
-        return this.compatibilityDestination[0];
-    }
-
-    /// <summary>
-    /// Adds values with the .NET runtime implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public T Runtime()
-    {
-        System.Numerics.Tensors.TensorPrimitives.Add<T>(this.x, this.y, this.runtimeDestination);
-        return this.runtimeDestination[0];
-    }
-}
-
-[GenericTypeArguments(typeof(byte))]
-[GenericTypeArguments(typeof(uint))]
-[GenericTypeArguments(typeof(int))]
-[GenericTypeArguments(typeof(float))]
-[GenericTypeArguments(typeof(double))]
-public class TensorPrimitivesRuntimeClampAssemblyComparison<T>
-    where T : unmanaged, INumber<T>
-{
-    private T[] compatibilityValues = null!;
-    private T[] runtimeValues = null!;
-    private T min;
-    private T max;
-
-    /// <summary>
-    /// Creates deterministic inputs for both implementations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.compatibilityValues = new T[2048];
-        this.runtimeValues = new T[2048];
-        this.min = T.CreateTruncating(64);
-        this.max = T.CreateTruncating(128);
-
-        for (int i = 0; i < this.compatibilityValues.Length; i++)
-        {
-            T value = T.CreateTruncating((i * 31) % 257);
-            this.compatibilityValues[i] = value;
-            this.runtimeValues[i] = value;
-        }
-    }
-
-    /// <summary>
-    /// Clamps values with the compatibility implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark(Baseline = true)]
-    public T Compatibility()
-    {
-        TensorPrimitives_.Clamp(this.compatibilityValues, this.min, this.max, this.compatibilityValues);
-        return this.compatibilityValues[0];
-    }
-
-    /// <summary>
-    /// Clamps values with the .NET runtime implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public T Runtime()
-    {
-        System.Numerics.Tensors.TensorPrimitives.Clamp(this.runtimeValues, this.min, this.max, this.runtimeValues);
-        return this.runtimeValues[0];
-    }
-}
-
-public class TensorPrimitivesRuntimeSingleScalarAssemblyComparison
-{
-    private readonly float scalar = -1F;
-    private float[] compatibilityValues = null!;
-    private float[] runtimeValues = null!;
-
-    /// <summary>
-    /// Creates equivalent stable inputs for both implementations.
-    /// </summary>
-    [GlobalSetup]
-    public void Setup()
-    {
-        this.compatibilityValues = new float[2048];
-        this.runtimeValues = new float[2048];
-
-        for (int i = 0; i < this.compatibilityValues.Length; i++)
-        {
-            float value = ((i * 17) % 251) + 1;
-            this.compatibilityValues[i] = value;
-            this.runtimeValues[i] = value;
-        }
-    }
-
-    /// <summary>
-    /// Divides values with the compatibility implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float CompatibilityDivide()
-    {
-        TensorPrimitives_.Divide(this.compatibilityValues, this.scalar, this.compatibilityValues);
-        return this.compatibilityValues[0];
-    }
-
-    /// <summary>
-    /// Divides values with the .NET runtime implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float RuntimeDivide()
-    {
-        System.Numerics.Tensors.TensorPrimitives.Divide(this.runtimeValues, this.scalar, this.runtimeValues);
-        return this.runtimeValues[0];
-    }
-
-    /// <summary>
-    /// Computes maximum values with the compatibility implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float CompatibilityMax()
-    {
-        TensorPrimitives_.Max(this.compatibilityValues, 0F, this.compatibilityValues);
-        return this.compatibilityValues[0];
-    }
-
-    /// <summary>
-    /// Computes maximum values with the .NET runtime implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float RuntimeMax()
-    {
-        System.Numerics.Tensors.TensorPrimitives.Max(this.runtimeValues, 0F, this.runtimeValues);
-        return this.runtimeValues[0];
-    }
-
-    /// <summary>
-    /// Multiplies values with the compatibility implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float CompatibilityMultiply()
-    {
-        TensorPrimitives_.Multiply(this.compatibilityValues, this.scalar, this.compatibilityValues);
-        return this.compatibilityValues[0];
-    }
-
-    /// <summary>
-    /// Multiplies values with the .NET runtime implementation.
-    /// </summary>
-    /// <returns>The first result, which keeps the writes observable to the benchmark harness.</returns>
-    [Benchmark]
-    public float RuntimeMultiply()
-    {
-        System.Numerics.Tensors.TensorPrimitives.Multiply(this.runtimeValues, this.scalar, this.runtimeValues);
-        return this.runtimeValues[0];
-    }
-}
-#endif

tests/ImageSharp.Benchmarks/General/PixelConversion/Vector4AffineTransform.cs

@@ -2,16 +2,13 @@
 // Licensed under the Six Labors Split License.
 
 using System.Numerics;
-using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
-using System.Runtime.Intrinsics;
 using BenchmarkDotNet.Attributes;
 using SixLabors.ImageSharp.PixelFormats.Utils;
 
 namespace SixLabors.ImageSharp.Benchmarks.General.PixelConversion;
 
 /// <summary>
-/// Compares operator-driven affine vector transforms with the duplicated traversals they replace.
+/// Measures the operator-driven affine vector transforms.
 /// </summary>
 [Config(typeof(Config.Short))]
 public class Vector4AffineTransform
@@ -21,7 +18,6 @@ public class Vector4AffineTransform
     private static readonly Vector4 Divisor = new(255F, 2F, 65535F, .5F);
 
     private Vector4[] current;
-    private Vector4[] baseline;
 
     /// <summary>
     /// Gets or sets the number of vectors transformed by each invocation.
@@ -30,7 +26,7 @@ public class Vector4AffineTransform
     public int Count { get; set; }
 
     /// <summary>
-    /// Creates identical non-uniform buffers for the current and baseline traversals.
+    /// Creates a non-uniform input buffer.
     /// </summary>
     [GlobalSetup]
     public void Setup()
@@ -41,167 +37,19 @@ public class Vector4AffineTransform
         {
             this.current[i] = new Vector4(i + .25F, i + .5F, i + .75F, i + 1F);
         }
-
-        this.baseline = [.. this.current];
     }
 
     /// <summary>
     /// Executes the operator-driven multiply-then-add traversal.
     /// </summary>
     [Benchmark]
-    public void CurrentMultiplyThenAdd()
+    public void MultiplyThenAdd()
         => Vector4Converters.MultiplyThenAdd(this.current, Multiplier, Offset);
 
-    /// <summary>
-    /// Executes the duplicated multiply-then-add traversal.
-    /// </summary>
-    [Benchmark(Baseline = true)]
-    public void BaselineMultiplyThenAdd()
-        => BaselineMultiplyThenAdd(this.baseline, Multiplier, Offset);
-
     /// <summary>
     /// Executes the operator-driven add-then-divide traversal.
     /// </summary>
     [Benchmark]
-    public void CurrentAddThenDivide()
+    public void AddThenDivide()
         => Vector4Converters.AddThenDivide(this.current, Offset, Divisor);
-
-    /// <summary>
-    /// Executes the duplicated add-then-divide traversal.
-    /// </summary>
-    [Benchmark]
-    public void BaselineAddThenDivide()
-        => BaselineAddThenDivide(this.baseline, Offset, Divisor);
-
-    /// <summary>
-    /// Retains the multiply-then-add traversal being replaced for direct measurement.
-    /// </summary>
-    /// <param name="vectors">The vectors to transform.</param>
-    /// <param name="multiplier">The component-wise multiplier.</param>
-    /// <param name="offset">The component-wise offset.</param>
-    internal static void BaselineMultiplyThenAdd(Span<Vector4> vectors, Vector4 multiplier, Vector4 offset)
-    {
-        ref Vector4 vectorBase = ref MemoryMarshal.GetReference(vectors);
-        int index = 0;
-
-        if (Vector512.IsHardwareAccelerated)
-        {
-            int vectorsPerVector = Vector512<float>.Count / Vector128<float>.Count;
-            Vector256<float> multiplier256 = Vector256.Create(multiplier.AsVector128(), multiplier.AsVector128());
-            Vector256<float> offset256 = Vector256.Create(offset.AsVector128(), offset.AsVector128());
-            Vector512<float> multiplier512 = Vector512.Create(multiplier256, multiplier256);
-            Vector512<float> offset512 = Vector512.Create(offset256, offset256);
-
-            for (; index <= vectors.Length - vectorsPerVector; index += vectorsPerVector)
-            {
-                ref Vector512<float> vector = ref Unsafe.As<Vector4, Vector512<float>>(
-                    ref Unsafe.Add(ref vectorBase, (uint)index));
-
-                vector = (vector * multiplier512) + offset512;
-            }
-        }
-
-        if (Vector256.IsHardwareAccelerated)
-        {
-            int vectorsPerVector = Vector256<float>.Count / Vector128<float>.Count;
-            Vector256<float> multiplier256 = Vector256.Create(multiplier.AsVector128(), multiplier.AsVector128());
-            Vector256<float> offset256 = Vector256.Create(offset.AsVector128(), offset.AsVector128());
-
-            for (; index <= vectors.Length - vectorsPerVector; index += vectorsPerVector)
-            {
-                ref Vector256<float> vector = ref Unsafe.As<Vector4, Vector256<float>>(
-                    ref Unsafe.Add(ref vectorBase, (uint)index));
-
-                vector = (vector * multiplier256) + offset256;
-            }
-        }
-
-        if (Vector128.IsHardwareAccelerated)
-        {
-            Vector128<float> multiplier128 = multiplier.AsVector128();
-            Vector128<float> offset128 = offset.AsVector128();
-
-            for (; index < vectors.Length; index++)
-            {
-                ref Vector128<float> vector = ref Unsafe.As<Vector4, Vector128<float>>(
-                    ref Unsafe.Add(ref vectorBase, (uint)index));
-
-                vector = (vector * multiplier128) + offset128;
-            }
-
-            return;
-        }
-
-        for (; index < vectors.Length; index++)
-        {
-            ref Vector4 vector = ref Unsafe.Add(ref vectorBase, (uint)index);
-            vector = (vector * multiplier) + offset;
-        }
-    }
-
-    /// <summary>
-    /// Retains the add-then-divide traversal being replaced for direct measurement.
-    /// </summary>
-    /// <param name="vectors">The vectors to transform.</param>
-    /// <param name="offset">The component-wise offset.</param>
-    /// <param name="divisor">The component-wise divisor.</param>
-    internal static void BaselineAddThenDivide(Span<Vector4> vectors, Vector4 offset, Vector4 divisor)
-    {
-        ref Vector4 vectorBase = ref MemoryMarshal.GetReference(vectors);
-        int index = 0;
-
-        if (Vector512.IsHardwareAccelerated)
-        {
-            int vectorsPerVector = Vector512<float>.Count / Vector128<float>.Count;
-            Vector256<float> offset256 = Vector256.Create(offset.AsVector128(), offset.AsVector128());
-            Vector256<float> divisor256 = Vector256.Create(divisor.AsVector128(), divisor.AsVector128());
-            Vector512<float> offset512 = Vector512.Create(offset256, offset256);
-            Vector512<float> divisor512 = Vector512.Create(divisor256, divisor256);
-
-            for (; index <= vectors.Length - vectorsPerVector; index += vectorsPerVector)
-            {
-                ref Vector512<float> vector = ref Unsafe.As<Vector4, Vector512<float>>(
-                    ref Unsafe.Add(ref vectorBase, (uint)index));
-
-                vector = (vector + offset512) / divisor512;
-            }
-        }
-
-        if (Vector256.IsHardwareAccelerated)
-        {
-            int vectorsPerVector = Vector256<float>.Count / Vector128<float>.Count;
-            Vector256<float> offset256 = Vector256.Create(offset.AsVector128(), offset.AsVector128());
-            Vector256<float> divisor256 = Vector256.Create(divisor.AsVector128(), divisor.AsVector128());
-
-            for (; index <= vectors.Length - vectorsPerVector; index += vectorsPerVector)
-            {
-                ref Vector256<float> vector = ref Unsafe.As<Vector4, Vector256<float>>(
-                    ref Unsafe.Add(ref vectorBase, (uint)index));
-
-                vector = (vector + offset256) / divisor256;
-            }
-        }
-
-        if (Vector128.IsHardwareAccelerated)
-        {
-            Vector128<float> offset128 = offset.AsVector128();
-            Vector128<float> divisor128 = divisor.AsVector128();
-
-            for (; index < vectors.Length; index++)
-            {
-                ref Vector128<float> vector = ref Unsafe.As<Vector4, Vector128<float>>(
-                    ref Unsafe.Add(ref vectorBase, (uint)index));
-
-                vector = (vector + offset128) / divisor128;
-            }
-
-            return;
-        }
-
-        for (; index < vectors.Length; index++)
-        {
-            ref Vector4 vector = ref Unsafe.Add(ref vectorBase, (uint)index);
-            vector = (vector + offset) / divisor;
-        }
-    }
 }

tests/ImageSharp.Benchmarks/General/PixelConversion/Vector4AffineTransformAssembly.cs

@@ -56,18 +56,4 @@ public class Vector4AffineTransformAssembly
     [Benchmark]
     public void AddThenDivide()
         => Vector4Converters.AddThenDivide(this.vectors, Offset, Divisor);
-
-    /// <summary>
-    /// Executes the multiply-then-add traversal being replaced for assembly comparison.
-    /// </summary>
-    [Benchmark]
-    public void BaselineMultiplyThenAdd()
-        => Vector4AffineTransform.BaselineMultiplyThenAdd(this.vectors, Multiplier, Offset);
-
-    /// <summary>
-    /// Executes the add-then-divide traversal being replaced for assembly comparison.
-    /// </summary>
-    [Benchmark]
-    public void BaselineAddThenDivide()
-        => Vector4AffineTransform.BaselineAddThenDivide(this.vectors, Offset, Divisor);
 }