From 9e7cf1611b8d5f50080715cf16f16b4773b42462 Mon Sep 17 00:00:00 2001
From: James Jackson-South <james_south@hotmail.com>
Date: Tue, 21 Feb 2023 21:41:20 +1000
Subject: [PATCH 1/3] Use MMShuffle consts everywhere and refactor explicit
 shuffle components to use them

---
 .../Helpers/Shuffle/IComponentShuffle.cs      |  54 ++--
 .../Common/Helpers/Shuffle/IPad3Shuffle4.cs   |  17 +-
 .../Common/Helpers/Shuffle/IShuffle3.cs       |   7 +-
 .../Common/Helpers/Shuffle/IShuffle4Slice3.cs |  15 +-
 .../Common/Helpers/SimdUtils.Shuffle.cs       | 271 +++++++++++++++++-
 .../Formats/Webp/Lossless/LosslessUtils.cs    |  58 ++--
 .../Formats/Webp/Lossy/Vp8Encoding.cs         |   9 +-
 .../Color/Bulk/ShuffleFloat4Channel.cs        |   5 +-
 .../Common/SimdUtilsTests.Shuffle.cs          |  50 ++--
 9 files changed, 369 insertions(+), 117 deletions(-)

diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs b/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs
index 45d6e6d13..b93202814 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs
@@ -5,6 +5,7 @@ using System.Buffers.Binary;
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
+using static SixLabors.ImageSharp.SimdUtils;
 
 // The JIT can detect and optimize rotation idioms ROTL (Rotate Left)
 // and ROTR (Rotate Right) emitting efficient CPU instructions:
@@ -18,9 +19,12 @@ namespace SixLabors.ImageSharp;
 internal interface IComponentShuffle
 {
     /// <summary>
-    /// Gets the shuffle control.
+    /// Shuffles then slices 8-bit integers within 128-bit lanes in <paramref name="source"/>
+    /// using the control and store the results in <paramref name="dest"/>.
     /// </summary>
-    byte Control { get; }
+    /// <param name="source">The source span of bytes.</param>
+    /// <param name="dest">The destination span of bytes.</param>
+    void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest);
 
     /// <summary>
     /// Shuffle 8-bit integers within 128-bit lanes in <paramref name="source"/>
@@ -58,11 +62,15 @@ internal readonly struct DefaultShuffle4 : IShuffle4
         this.p2 = p2;
         this.p1 = p1;
         this.p0 = p0;
-        this.Control = SimdUtils.Shuffle.MmShuffle(p3, p2, p1, p0);
+        this.Control = Shuffle.MmShuffle(p3, p2, p1, p0);
     }
 
     public byte Control { get; }
 
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Shuffle4Reduce(ref source, ref dest, this.Control);
+
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
     {
@@ -86,11 +94,9 @@ internal readonly struct DefaultShuffle4 : IShuffle4
 
 internal readonly struct WXYZShuffle4 : IShuffle4
 {
-    public byte Control
-    {
-        [MethodImpl(InliningOptions.ShortMethod)]
-        get => SimdUtils.Shuffle.MmShuffle(2, 1, 0, 3);
-    }
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Shuffle4Reduce(ref source, ref dest, Shuffle.MMShuffle2103);
 
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
@@ -112,11 +118,9 @@ internal readonly struct WXYZShuffle4 : IShuffle4
 
 internal readonly struct WZYXShuffle4 : IShuffle4
 {
-    public byte Control
-    {
-        [MethodImpl(InliningOptions.ShortMethod)]
-        get => SimdUtils.Shuffle.MmShuffle(0, 1, 2, 3);
-    }
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Shuffle4Reduce(ref source, ref dest, Shuffle.MMShuffle0123);
 
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
@@ -138,11 +142,9 @@ internal readonly struct WZYXShuffle4 : IShuffle4
 
 internal readonly struct YZWXShuffle4 : IShuffle4
 {
-    public byte Control
-    {
-        [MethodImpl(InliningOptions.ShortMethod)]
-        get => SimdUtils.Shuffle.MmShuffle(0, 3, 2, 1);
-    }
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Shuffle4Reduce(ref source, ref dest, Shuffle.MMShuffle0321);
 
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
@@ -164,11 +166,9 @@ internal readonly struct YZWXShuffle4 : IShuffle4
 
 internal readonly struct ZYXWShuffle4 : IShuffle4
 {
-    public byte Control
-    {
-        [MethodImpl(InliningOptions.ShortMethod)]
-        get => SimdUtils.Shuffle.MmShuffle(3, 0, 1, 2);
-    }
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Shuffle4Reduce(ref source, ref dest, Shuffle.MMShuffle3012);
 
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
@@ -197,11 +197,9 @@ internal readonly struct ZYXWShuffle4 : IShuffle4
 
 internal readonly struct XWZYShuffle4 : IShuffle4
 {
-    public byte Control
-    {
-        [MethodImpl(InliningOptions.ShortMethod)]
-        get => SimdUtils.Shuffle.MmShuffle(1, 2, 3, 0);
-    }
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Shuffle4Reduce(ref source, ref dest, Shuffle.MMShuffle1230);
 
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs b/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs
index 76cffd82b..7f99b3a08 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs
@@ -3,6 +3,7 @@
 
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
+using static SixLabors.ImageSharp.SimdUtils;
 
 namespace SixLabors.ImageSharp;
 
@@ -29,11 +30,15 @@ internal readonly struct DefaultPad3Shuffle4 : IPad3Shuffle4
         this.p2 = p2;
         this.p1 = p1;
         this.p0 = p0;
-        this.Control = SimdUtils.Shuffle.MmShuffle(p3, p2, p1, p0);
+        this.Control = Shuffle.MmShuffle(p3, p2, p1, p0);
     }
 
     public byte Control { get; }
 
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Pad3Shuffle4Reduce(ref source, ref dest, this.Control);
+
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
     {
@@ -51,7 +56,7 @@ internal readonly struct DefaultPad3Shuffle4 : IPad3Shuffle4
 
         for (int i = 0, j = 0; i < source.Length; i += 3, j += 4)
         {
-            ref var s = ref Unsafe.Add(ref sBase, i);
+            ref byte s = ref Unsafe.Add(ref sBase, i);
             tu = Unsafe.As<byte, uint>(ref s) | 0xFF000000;
 
             Unsafe.Add(ref dBase, j) = Unsafe.Add(ref t, p0);
@@ -64,11 +69,9 @@ internal readonly struct DefaultPad3Shuffle4 : IPad3Shuffle4
 
 internal readonly struct XYZWPad3Shuffle4 : IPad3Shuffle4
 {
-    public byte Control
-    {
-        [MethodImpl(InliningOptions.ShortMethod)]
-        get => SimdUtils.Shuffle.MmShuffle(3, 2, 1, 0);
-    }
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Pad3Shuffle4Reduce(ref source, ref dest, Shuffle.MMShuffle3210);
 
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs
index 9bee9d15e..edbd2cd5e 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs
@@ -3,6 +3,7 @@
 
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
+using static SixLabors.ImageSharp.SimdUtils;
 
 namespace SixLabors.ImageSharp;
 
@@ -26,11 +27,15 @@ internal readonly struct DefaultShuffle3 : IShuffle3
         this.p2 = p2;
         this.p1 = p1;
         this.p0 = p0;
-        this.Control = SimdUtils.Shuffle.MmShuffle(3, p2, p1, p0);
+        this.Control = Shuffle.MmShuffle(3, p2, p1, p0);
     }
 
     public byte Control { get; }
 
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Shuffle3Reduce(ref source, ref dest, this.Control);
+
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
     {
diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs
index 90b77b568..2179a085b 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs
@@ -3,6 +3,7 @@
 
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
+using static SixLabors.ImageSharp.SimdUtils;
 
 namespace SixLabors.ImageSharp;
 
@@ -27,11 +28,15 @@ internal readonly struct DefaultShuffle4Slice3 : IShuffle4Slice3
         this.p2 = p2;
         this.p1 = p1;
         this.p0 = p0;
-        this.Control = SimdUtils.Shuffle.MmShuffle(p3, p2, p1, p0);
+        this.Control = Shuffle.MmShuffle(p3, p2, p1, p0);
     }
 
     public byte Control { get; }
 
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Shuffle4Slice3Reduce(ref source, ref dest, this.Control);
+
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
     {
@@ -53,11 +58,9 @@ internal readonly struct DefaultShuffle4Slice3 : IShuffle4Slice3
 
 internal readonly struct XYZWShuffle4Slice3 : IShuffle4Slice3
 {
-    public byte Control
-    {
-        [MethodImpl(InliningOptions.ShortMethod)]
-        get => SimdUtils.Shuffle.MmShuffle(3, 2, 1, 0);
-    }
+    [MethodImpl(InliningOptions.ShortMethod)]
+    public void ShuffleReduce(ref ReadOnlySpan<byte> source, ref Span<byte> dest)
+        => HwIntrinsics.Shuffle4Slice3Reduce(ref source, ref dest, Shuffle.MMShuffle3210);
 
     [MethodImpl(InliningOptions.ShortMethod)]
     public void RunFallbackShuffle(ReadOnlySpan<byte> source, Span<byte> dest)
diff --git a/src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs b/src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs
index b91dc2fad..4ff5f3d5b 100644
--- a/src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs
+++ b/src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs
@@ -37,6 +37,7 @@ internal static partial class SimdUtils
     /// Shuffle 8-bit integers within 128-bit lanes in <paramref name="source"/>
     /// using the control and store the results in <paramref name="dest"/>.
     /// </summary>
+    /// <typeparam name="TShuffle">The type of shuffle struct.</typeparam>
     /// <param name="source">The source span of bytes.</param>
     /// <param name="dest">The destination span of bytes.</param>
     /// <param name="shuffle">The type of shuffle to perform.</param>
@@ -49,7 +50,7 @@ internal static partial class SimdUtils
     {
         VerifyShuffle4SpanInput(source, dest);
 
-        HwIntrinsics.Shuffle4Reduce(ref source, ref dest, shuffle.Control);
+        shuffle.ShuffleReduce(ref source, ref dest);
 
         // Deal with the remainder:
         if (source.Length > 0)
@@ -62,6 +63,7 @@ internal static partial class SimdUtils
     /// Shuffle 8-bit integer triplets within 128-bit lanes in <paramref name="source"/>
     /// using the control and store the results in <paramref name="dest"/>.
     /// </summary>
+    /// <typeparam name="TShuffle">The type of shuffle struct.</typeparam>
     /// <param name="source">The source span of bytes.</param>
     /// <param name="dest">The destination span of bytes.</param>
     /// <param name="shuffle">The type of shuffle to perform.</param>
@@ -75,7 +77,7 @@ internal static partial class SimdUtils
         // Source length should be smaller than dest length, and divisible by 3.
         VerifyShuffle3SpanInput(source, dest);
 
-        HwIntrinsics.Shuffle3Reduce(ref source, ref dest, shuffle.Control);
+        shuffle.ShuffleReduce(ref source, ref dest);
 
         // Deal with the remainder:
         if (source.Length > 0)
@@ -88,6 +90,7 @@ internal static partial class SimdUtils
     /// Pads then shuffles 8-bit integers within 128-bit lanes in <paramref name="source"/>
     /// using the control and store the results in <paramref name="dest"/>.
     /// </summary>
+    /// <typeparam name="TShuffle">The type of shuffle struct.</typeparam>
     /// <param name="source">The source span of bytes.</param>
     /// <param name="dest">The destination span of bytes.</param>
     /// <param name="shuffle">The type of shuffle to perform.</param>
@@ -100,7 +103,7 @@ internal static partial class SimdUtils
     {
         VerifyPad3Shuffle4SpanInput(source, dest);
 
-        HwIntrinsics.Pad3Shuffle4Reduce(ref source, ref dest, shuffle.Control);
+        shuffle.ShuffleReduce(ref source, ref dest);
 
         // Deal with the remainder:
         if (source.Length > 0)
@@ -113,6 +116,7 @@ internal static partial class SimdUtils
     /// Shuffles then slices 8-bit integers within 128-bit lanes in <paramref name="source"/>
     /// using the control and store the results in <paramref name="dest"/>.
     /// </summary>
+    /// <typeparam name="TShuffle">The type of shuffle struct.</typeparam>
     /// <param name="source">The source span of bytes.</param>
     /// <param name="dest">The destination span of bytes.</param>
     /// <param name="shuffle">The type of shuffle to perform.</param>
@@ -125,7 +129,7 @@ internal static partial class SimdUtils
     {
         VerifyShuffle4Slice3SpanInput(source, dest);
 
-        HwIntrinsics.Shuffle4Slice3Reduce(ref source, ref dest, shuffle.Control);
+        shuffle.ShuffleReduce(ref source, ref dest);
 
         // Deal with the remainder:
         if (source.Length > 0)
@@ -153,7 +157,7 @@ internal static partial class SimdUtils
     }
 
     [Conditional("DEBUG")]
-    private static void VerifyShuffle4SpanInput<T>(ReadOnlySpan<T> source, Span<T> dest)
+    internal static void VerifyShuffle4SpanInput<T>(ReadOnlySpan<T> source, Span<T> dest)
         where T : struct
     {
         DebugGuard.IsTrue(
@@ -222,6 +226,263 @@ internal static partial class SimdUtils
 
     public static class Shuffle
     {
+        public const byte MMShuffle0000 = 0b00000000;
+        public const byte MMShuffle0001 = 0b00000001;
+        public const byte MMShuffle0002 = 0b00000010;
+        public const byte MMShuffle0003 = 0b00000011;
+        public const byte MMShuffle0010 = 0b00000100;
+        public const byte MMShuffle0011 = 0b00000101;
+        public const byte MMShuffle0012 = 0b00000110;
+        public const byte MMShuffle0013 = 0b00000111;
+        public const byte MMShuffle0020 = 0b00001000;
+        public const byte MMShuffle0021 = 0b00001001;
+        public const byte MMShuffle0022 = 0b00001010;
+        public const byte MMShuffle0023 = 0b00001011;
+        public const byte MMShuffle0030 = 0b00001100;
+        public const byte MMShuffle0031 = 0b00001101;
+        public const byte MMShuffle0032 = 0b00001110;
+        public const byte MMShuffle0033 = 0b00001111;
+        public const byte MMShuffle0100 = 0b00010000;
+        public const byte MMShuffle0101 = 0b00010001;
+        public const byte MMShuffle0102 = 0b00010010;
+        public const byte MMShuffle0103 = 0b00010011;
+        public const byte MMShuffle0110 = 0b00010100;
+        public const byte MMShuffle0111 = 0b00010101;
+        public const byte MMShuffle0112 = 0b00010110;
+        public const byte MMShuffle0113 = 0b00010111;
+        public const byte MMShuffle0120 = 0b00011000;
+        public const byte MMShuffle0121 = 0b00011001;
+        public const byte MMShuffle0122 = 0b00011010;
+        public const byte MMShuffle0123 = 0b00011011;
+        public const byte MMShuffle0130 = 0b00011100;
+        public const byte MMShuffle0131 = 0b00011101;
+        public const byte MMShuffle0132 = 0b00011110;
+        public const byte MMShuffle0133 = 0b00011111;
+        public const byte MMShuffle0200 = 0b00100000;
+        public const byte MMShuffle0201 = 0b00100001;
+        public const byte MMShuffle0202 = 0b00100010;
+        public const byte MMShuffle0203 = 0b00100011;
+        public const byte MMShuffle0210 = 0b00100100;
+        public const byte MMShuffle0211 = 0b00100101;
+        public const byte MMShuffle0212 = 0b00100110;
+        public const byte MMShuffle0213 = 0b00100111;
+        public const byte MMShuffle0220 = 0b00101000;
+        public const byte MMShuffle0221 = 0b00101001;
+        public const byte MMShuffle0222 = 0b00101010;
+        public const byte MMShuffle0223 = 0b00101011;
+        public const byte MMShuffle0230 = 0b00101100;
+        public const byte MMShuffle0231 = 0b00101101;
+        public const byte MMShuffle0232 = 0b00101110;
+        public const byte MMShuffle0233 = 0b00101111;
+        public const byte MMShuffle0300 = 0b00110000;
+        public const byte MMShuffle0301 = 0b00110001;
+        public const byte MMShuffle0302 = 0b00110010;
+        public const byte MMShuffle0303 = 0b00110011;
+        public const byte MMShuffle0310 = 0b00110100;
+        public const byte MMShuffle0311 = 0b00110101;
+        public const byte MMShuffle0312 = 0b00110110;
+        public const byte MMShuffle0313 = 0b00110111;
+        public const byte MMShuffle0320 = 0b00111000;
+        public const byte MMShuffle0321 = 0b00111001;
+        public const byte MMShuffle0322 = 0b00111010;
+        public const byte MMShuffle0323 = 0b00111011;
+        public const byte MMShuffle0330 = 0b00111100;
+        public const byte MMShuffle0331 = 0b00111101;
+        public const byte MMShuffle0332 = 0b00111110;
+        public const byte MMShuffle0333 = 0b00111111;
+        public const byte MMShuffle1000 = 0b01000000;
+        public const byte MMShuffle1001 = 0b01000001;
+        public const byte MMShuffle1002 = 0b01000010;
+        public const byte MMShuffle1003 = 0b01000011;
+        public const byte MMShuffle1010 = 0b01000100;
+        public const byte MMShuffle1011 = 0b01000101;
+        public const byte MMShuffle1012 = 0b01000110;
+        public const byte MMShuffle1013 = 0b01000111;
+        public const byte MMShuffle1020 = 0b01001000;
+        public const byte MMShuffle1021 = 0b01001001;
+        public const byte MMShuffle1022 = 0b01001010;
+        public const byte MMShuffle1023 = 0b01001011;
+        public const byte MMShuffle1030 = 0b01001100;
+        public const byte MMShuffle1031 = 0b01001101;
+        public const byte MMShuffle1032 = 0b01001110;
+        public const byte MMShuffle1033 = 0b01001111;
+        public const byte MMShuffle1100 = 0b01010000;
+        public const byte MMShuffle1101 = 0b01010001;
+        public const byte MMShuffle1102 = 0b01010010;
+        public const byte MMShuffle1103 = 0b01010011;
+        public const byte MMShuffle1110 = 0b01010100;
+        public const byte MMShuffle1111 = 0b01010101;
+        public const byte MMShuffle1112 = 0b01010110;
+        public const byte MMShuffle1113 = 0b01010111;
+        public const byte MMShuffle1120 = 0b01011000;
+        public const byte MMShuffle1121 = 0b01011001;
+        public const byte MMShuffle1122 = 0b01011010;
+        public const byte MMShuffle1123 = 0b01011011;
+        public const byte MMShuffle1130 = 0b01011100;
+        public const byte MMShuffle1131 = 0b01011101;
+        public const byte MMShuffle1132 = 0b01011110;
+        public const byte MMShuffle1133 = 0b01011111;
+        public const byte MMShuffle1200 = 0b01100000;
+        public const byte MMShuffle1201 = 0b01100001;
+        public const byte MMShuffle1202 = 0b01100010;
+        public const byte MMShuffle1203 = 0b01100011;
+        public const byte MMShuffle1210 = 0b01100100;
+        public const byte MMShuffle1211 = 0b01100101;
+        public const byte MMShuffle1212 = 0b01100110;
+        public const byte MMShuffle1213 = 0b01100111;
+        public const byte MMShuffle1220 = 0b01101000;
+        public const byte MMShuffle1221 = 0b01101001;
+        public const byte MMShuffle1222 = 0b01101010;
+        public const byte MMShuffle1223 = 0b01101011;
+        public const byte MMShuffle1230 = 0b01101100;
+        public const byte MMShuffle1231 = 0b01101101;
+        public const byte MMShuffle1232 = 0b01101110;
+        public const byte MMShuffle1233 = 0b01101111;
+        public const byte MMShuffle1300 = 0b01110000;
+        public const byte MMShuffle1301 = 0b01110001;
+        public const byte MMShuffle1302 = 0b01110010;
+        public const byte MMShuffle1303 = 0b01110011;
+        public const byte MMShuffle1310 = 0b01110100;
+        public const byte MMShuffle1311 = 0b01110101;
+        public const byte MMShuffle1312 = 0b01110110;
+        public const byte MMShuffle1313 = 0b01110111;
+        public const byte MMShuffle1320 = 0b01111000;
+        public const byte MMShuffle1321 = 0b01111001;
+        public const byte MMShuffle1322 = 0b01111010;
+        public const byte MMShuffle1323 = 0b01111011;
+        public const byte MMShuffle1330 = 0b01111100;
+        public const byte MMShuffle1331 = 0b01111101;
+        public const byte MMShuffle1332 = 0b01111110;
+        public const byte MMShuffle1333 = 0b01111111;
+        public const byte MMShuffle2000 = 0b10000000;
+        public const byte MMShuffle2001 = 0b10000001;
+        public const byte MMShuffle2002 = 0b10000010;
+        public const byte MMShuffle2003 = 0b10000011;
+        public const byte MMShuffle2010 = 0b10000100;
+        public const byte MMShuffle2011 = 0b10000101;
+        public const byte MMShuffle2012 = 0b10000110;
+        public const byte MMShuffle2013 = 0b10000111;
+        public const byte MMShuffle2020 = 0b10001000;
+        public const byte MMShuffle2021 = 0b10001001;
+        public const byte MMShuffle2022 = 0b10001010;
+        public const byte MMShuffle2023 = 0b10001011;
+        public const byte MMShuffle2030 = 0b10001100;
+        public const byte MMShuffle2031 = 0b10001101;
+        public const byte MMShuffle2032 = 0b10001110;
+        public const byte MMShuffle2033 = 0b10001111;
+        public const byte MMShuffle2100 = 0b10010000;
+        public const byte MMShuffle2101 = 0b10010001;
+        public const byte MMShuffle2102 = 0b10010010;
+        public const byte MMShuffle2103 = 0b10010011;
+        public const byte MMShuffle2110 = 0b10010100;
+        public const byte MMShuffle2111 = 0b10010101;
+        public const byte MMShuffle2112 = 0b10010110;
+        public const byte MMShuffle2113 = 0b10010111;
+        public const byte MMShuffle2120 = 0b10011000;
+        public const byte MMShuffle2121 = 0b10011001;
+        public const byte MMShuffle2122 = 0b10011010;
+        public const byte MMShuffle2123 = 0b10011011;
+        public const byte MMShuffle2130 = 0b10011100;
+        public const byte MMShuffle2131 = 0b10011101;
+        public const byte MMShuffle2132 = 0b10011110;
+        public const byte MMShuffle2133 = 0b10011111;
+        public const byte MMShuffle2200 = 0b10100000;
+        public const byte MMShuffle2201 = 0b10100001;
+        public const byte MMShuffle2202 = 0b10100010;
+        public const byte MMShuffle2203 = 0b10100011;
+        public const byte MMShuffle2210 = 0b10100100;
+        public const byte MMShuffle2211 = 0b10100101;
+        public const byte MMShuffle2212 = 0b10100110;
+        public const byte MMShuffle2213 = 0b10100111;
+        public const byte MMShuffle2220 = 0b10101000;
+        public const byte MMShuffle2221 = 0b10101001;
+        public const byte MMShuffle2222 = 0b10101010;
+        public const byte MMShuffle2223 = 0b10101011;
+        public const byte MMShuffle2230 = 0b10101100;
+        public const byte MMShuffle2231 = 0b10101101;
+        public const byte MMShuffle2232 = 0b10101110;
+        public const byte MMShuffle2233 = 0b10101111;
+        public const byte MMShuffle2300 = 0b10110000;
+        public const byte MMShuffle2301 = 0b10110001;
+        public const byte MMShuffle2302 = 0b10110010;
+        public const byte MMShuffle2303 = 0b10110011;
+        public const byte MMShuffle2310 = 0b10110100;
+        public const byte MMShuffle2311 = 0b10110101;
+        public const byte MMShuffle2312 = 0b10110110;
+        public const byte MMShuffle2313 = 0b10110111;
+        public const byte MMShuffle2320 = 0b10111000;
+        public const byte MMShuffle2321 = 0b10111001;
+        public const byte MMShuffle2322 = 0b10111010;
+        public const byte MMShuffle2323 = 0b10111011;
+        public const byte MMShuffle2330 = 0b10111100;
+        public const byte MMShuffle2331 = 0b10111101;
+        public const byte MMShuffle2332 = 0b10111110;
+        public const byte MMShuffle2333 = 0b10111111;
+        public const byte MMShuffle3000 = 0b11000000;
+        public const byte MMShuffle3001 = 0b11000001;
+        public const byte MMShuffle3002 = 0b11000010;
+        public const byte MMShuffle3003 = 0b11000011;
+        public const byte MMShuffle3010 = 0b11000100;
+        public const byte MMShuffle3011 = 0b11000101;
+        public const byte MMShuffle3012 = 0b11000110;
+        public const byte MMShuffle3013 = 0b11000111;
+        public const byte MMShuffle3020 = 0b11001000;
+        public const byte MMShuffle3021 = 0b11001001;
+        public const byte MMShuffle3022 = 0b11001010;
+        public const byte MMShuffle3023 = 0b11001011;
+        public const byte MMShuffle3030 = 0b11001100;
+        public const byte MMShuffle3031 = 0b11001101;
+        public const byte MMShuffle3032 = 0b11001110;
+        public const byte MMShuffle3033 = 0b11001111;
+        public const byte MMShuffle3100 = 0b11010000;
+        public const byte MMShuffle3101 = 0b11010001;
+        public const byte MMShuffle3102 = 0b11010010;
+        public const byte MMShuffle3103 = 0b11010011;
+        public const byte MMShuffle3110 = 0b11010100;
+        public const byte MMShuffle3111 = 0b11010101;
+        public const byte MMShuffle3112 = 0b11010110;
+        public const byte MMShuffle3113 = 0b11010111;
+        public const byte MMShuffle3120 = 0b11011000;
+        public const byte MMShuffle3121 = 0b11011001;
+        public const byte MMShuffle3122 = 0b11011010;
+        public const byte MMShuffle3123 = 0b11011011;
+        public const byte MMShuffle3130 = 0b11011100;
+        public const byte MMShuffle3131 = 0b11011101;
+        public const byte MMShuffle3132 = 0b11011110;
+        public const byte MMShuffle3133 = 0b11011111;
+        public const byte MMShuffle3200 = 0b11100000;
+        public const byte MMShuffle3201 = 0b11100001;
+        public const byte MMShuffle3202 = 0b11100010;
+        public const byte MMShuffle3203 = 0b11100011;
+        public const byte MMShuffle3210 = 0b11100100;
+        public const byte MMShuffle3211 = 0b11100101;
+        public const byte MMShuffle3212 = 0b11100110;
+        public const byte MMShuffle3213 = 0b11100111;
+        public const byte MMShuffle3220 = 0b11101000;
+        public const byte MMShuffle3221 = 0b11101001;
+        public const byte MMShuffle3222 = 0b11101010;
+        public const byte MMShuffle3223 = 0b11101011;
+        public const byte MMShuffle3230 = 0b11101100;
+        public const byte MMShuffle3231 = 0b11101101;
+        public const byte MMShuffle3232 = 0b11101110;
+        public const byte MMShuffle3233 = 0b11101111;
+        public const byte MMShuffle3300 = 0b11110000;
+        public const byte MMShuffle3301 = 0b11110001;
+        public const byte MMShuffle3302 = 0b11110010;
+        public const byte MMShuffle3303 = 0b11110011;
+        public const byte MMShuffle3310 = 0b11110100;
+        public const byte MMShuffle3311 = 0b11110101;
+        public const byte MMShuffle3312 = 0b11110110;
+        public const byte MMShuffle3313 = 0b11110111;
+        public const byte MMShuffle3320 = 0b11111000;
+        public const byte MMShuffle3321 = 0b11111001;
+        public const byte MMShuffle3322 = 0b11111010;
+        public const byte MMShuffle3323 = 0b11111011;
+        public const byte MMShuffle3330 = 0b11111100;
+        public const byte MMShuffle3331 = 0b11111101;
+        public const byte MMShuffle3332 = 0b11111110;
+        public const byte MMShuffle3333 = 0b11111111;
+
         [MethodImpl(InliningOptions.ShortMethod)]
         public static byte MmShuffle(byte p3, byte p2, byte p1, byte p0)
             => (byte)((p3 << 6) | (p2 << 4) | (p1 << 2) | p0);
diff --git a/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs b/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
index 5d9c20709..ac92ce6a6 100644
--- a/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
+++ b/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
@@ -96,7 +96,7 @@ internal static unsafe class LosslessUtils
     {
         if (Avx2.IsSupported)
         {
-            var addGreenToBlueAndRedMaskAvx2 = Vector256.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255, 17, 255, 17, 255, 21, 255, 21, 255, 25, 255, 25, 255, 29, 255, 29, 255);
+            Vector256<byte> addGreenToBlueAndRedMaskAvx2 = Vector256.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255, 17, 255, 17, 255, 21, 255, 21, 255, 25, 255, 25, 255, 29, 255, 29, 255);
             int numPixels = pixelData.Length;
             nint i;
             for (i = 0; i <= numPixels - 8; i += 8)
@@ -115,7 +115,7 @@ internal static unsafe class LosslessUtils
         }
         else if (Ssse3.IsSupported)
         {
-            var addGreenToBlueAndRedMaskSsse3 = Vector128.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255);
+            Vector128<byte> addGreenToBlueAndRedMaskSsse3 = Vector128.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255);
             int numPixels = pixelData.Length;
             nint i;
             for (i = 0; i <= numPixels - 4; i += 4)
@@ -138,13 +138,11 @@ internal static unsafe class LosslessUtils
             nint i;
             for (i = 0; i <= numPixels - 4; i += 4)
             {
-                const byte mmShuffle_2200 = 0b_10_10_00_00;
-
                 ref uint pos = ref Unsafe.Add(ref MemoryMarshal.GetReference(pixelData), i);
                 Vector128<byte> input = Unsafe.As<uint, Vector128<uint>>(ref pos).AsByte();
                 Vector128<ushort> a = Sse2.ShiftRightLogical(input.AsUInt16(), 8); // 0 a 0 g
-                Vector128<ushort> b = Sse2.ShuffleLow(a, mmShuffle_2200);
-                Vector128<ushort> c = Sse2.ShuffleHigh(b, mmShuffle_2200); // 0g0g
+                Vector128<ushort> b = Sse2.ShuffleLow(a, SimdUtils.Shuffle.MMShuffle2200);
+                Vector128<ushort> c = Sse2.ShuffleHigh(b, SimdUtils.Shuffle.MMShuffle2200); // 0g0g
                 Vector128<byte> output = Sse2.Add(input.AsByte(), c.AsByte());
                 Unsafe.As<uint, Vector128<uint>>(ref pos) = output.AsUInt32();
             }
@@ -178,7 +176,7 @@ internal static unsafe class LosslessUtils
     {
         if (Avx2.IsSupported)
         {
-            var subtractGreenFromBlueAndRedMaskAvx2 = Vector256.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255, 17, 255, 17, 255, 21, 255, 21, 255, 25, 255, 25, 255, 29, 255, 29, 255);
+            Vector256<byte> subtractGreenFromBlueAndRedMaskAvx2 = Vector256.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255, 17, 255, 17, 255, 21, 255, 21, 255, 25, 255, 25, 255, 29, 255, 29, 255);
             int numPixels = pixelData.Length;
             nint i;
             for (i = 0; i <= numPixels - 8; i += 8)
@@ -197,7 +195,7 @@ internal static unsafe class LosslessUtils
         }
         else if (Ssse3.IsSupported)
         {
-            var subtractGreenFromBlueAndRedMaskSsse3 = Vector128.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255);
+            Vector128<byte> subtractGreenFromBlueAndRedMaskSsse3 = Vector128.Create(1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255);
             int numPixels = pixelData.Length;
             nint i;
             for (i = 0; i <= numPixels - 4; i += 4)
@@ -220,13 +218,11 @@ internal static unsafe class LosslessUtils
             nint i;
             for (i = 0; i <= numPixels - 4; i += 4)
             {
-                const byte mmShuffle_2200 = 0b_10_10_00_00;
-
                 ref uint pos = ref Unsafe.Add(ref MemoryMarshal.GetReference(pixelData), i);
                 Vector128<byte> input = Unsafe.As<uint, Vector128<uint>>(ref pos).AsByte();
                 Vector128<ushort> a = Sse2.ShiftRightLogical(input.AsUInt16(), 8); // 0 a 0 g
-                Vector128<ushort> b = Sse2.ShuffleLow(a, mmShuffle_2200);
-                Vector128<ushort> c = Sse2.ShuffleHigh(b, mmShuffle_2200); // 0g0g
+                Vector128<ushort> b = Sse2.ShuffleLow(a, SimdUtils.Shuffle.MMShuffle2200);
+                Vector128<ushort> c = Sse2.ShuffleHigh(b, SimdUtils.Shuffle.MMShuffle2200); // 0g0g
                 Vector128<byte> output = Sse2.Subtract(input.AsByte(), c.AsByte());
                 Unsafe.As<uint, Vector128<uint>>(ref pos) = output.AsUInt32();
             }
@@ -334,7 +330,7 @@ internal static unsafe class LosslessUtils
         while (y < yEnd)
         {
             int predRowIdx = predRowIdxStart;
-            var m = default(Vp8LMultipliers);
+            Vp8LMultipliers m = default;
             int srcSafeEnd = pixelPos + safeWidth;
             int srcEnd = pixelPos + width;
             while (pixelPos < srcSafeEnd)
@@ -371,21 +367,19 @@ internal static unsafe class LosslessUtils
     {
         if (Avx2.IsSupported && numPixels >= 8)
         {
-            var transformColorAlphaGreenMask256 = Vector256.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
-            var transformColorRedBlueMask256 = Vector256.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0);
+            Vector256<byte> transformColorAlphaGreenMask256 = Vector256.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
+            Vector256<byte> transformColorRedBlueMask256 = Vector256.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0);
             Vector256<int> multsrb = MkCst32(Cst5b(m.GreenToRed), Cst5b(m.GreenToBlue));
             Vector256<int> multsb2 = MkCst32(Cst5b(m.RedToBlue), 0);
 
             nint idx;
             for (idx = 0; idx <= numPixels - 8; idx += 8)
             {
-                const byte mmShuffle_2200 = 0b_10_10_00_00;
-
                 ref uint pos = ref Unsafe.Add(ref MemoryMarshal.GetReference(pixelData), idx);
                 Vector256<uint> input = Unsafe.As<uint, Vector256<uint>>(ref pos);
                 Vector256<byte> a = Avx2.And(input.AsByte(), transformColorAlphaGreenMask256);
-                Vector256<short> b = Avx2.ShuffleLow(a.AsInt16(), mmShuffle_2200);
-                Vector256<short> c = Avx2.ShuffleHigh(b.AsInt16(), mmShuffle_2200);
+                Vector256<short> b = Avx2.ShuffleLow(a.AsInt16(), SimdUtils.Shuffle.MMShuffle2200);
+                Vector256<short> c = Avx2.ShuffleHigh(b.AsInt16(), SimdUtils.Shuffle.MMShuffle2200);
                 Vector256<short> d = Avx2.MultiplyHigh(c.AsInt16(), multsrb.AsInt16());
                 Vector256<short> e = Avx2.ShiftLeftLogical(input.AsInt16(), 8);
                 Vector256<short> f = Avx2.MultiplyHigh(e.AsInt16(), multsb2.AsInt16());
@@ -403,20 +397,18 @@ internal static unsafe class LosslessUtils
         }
         else if (Sse2.IsSupported)
         {
-            var transformColorAlphaGreenMask = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
-            var transformColorRedBlueMask = Vector128.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0);
+            Vector128<byte> transformColorAlphaGreenMask = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
+            Vector128<byte> transformColorRedBlueMask = Vector128.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0);
             Vector128<int> multsrb = MkCst16(Cst5b(m.GreenToRed), Cst5b(m.GreenToBlue));
             Vector128<int> multsb2 = MkCst16(Cst5b(m.RedToBlue), 0);
             nint idx;
             for (idx = 0; idx <= numPixels - 4; idx += 4)
             {
-                const byte mmShuffle_2200 = 0b_10_10_00_00;
-
                 ref uint pos = ref Unsafe.Add(ref MemoryMarshal.GetReference(pixelData), idx);
                 Vector128<uint> input = Unsafe.As<uint, Vector128<uint>>(ref pos);
                 Vector128<byte> a = Sse2.And(input.AsByte(), transformColorAlphaGreenMask);
-                Vector128<short> b = Sse2.ShuffleLow(a.AsInt16(), mmShuffle_2200);
-                Vector128<short> c = Sse2.ShuffleHigh(b.AsInt16(), mmShuffle_2200);
+                Vector128<short> b = Sse2.ShuffleLow(a.AsInt16(), SimdUtils.Shuffle.MMShuffle2200);
+                Vector128<short> c = Sse2.ShuffleHigh(b.AsInt16(), SimdUtils.Shuffle.MMShuffle2200);
                 Vector128<short> d = Sse2.MultiplyHigh(c.AsInt16(), multsrb.AsInt16());
                 Vector128<short> e = Sse2.ShiftLeftLogical(input.AsInt16(), 8);
                 Vector128<short> f = Sse2.MultiplyHigh(e.AsInt16(), multsb2.AsInt16());
@@ -465,19 +457,17 @@ internal static unsafe class LosslessUtils
     {
         if (Avx2.IsSupported && pixelData.Length >= 8)
         {
-            var transformColorInverseAlphaGreenMask256 = Vector256.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
+            Vector256<byte> transformColorInverseAlphaGreenMask256 = Vector256.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
             Vector256<int> multsrb = MkCst32(Cst5b(m.GreenToRed), Cst5b(m.GreenToBlue));
             Vector256<int> multsb2 = MkCst32(Cst5b(m.RedToBlue), 0);
             nint idx;
             for (idx = 0; idx <= pixelData.Length - 8; idx += 8)
             {
-                const byte mmShuffle_2200 = 0b_10_10_00_00;
-
                 ref uint pos = ref Unsafe.Add(ref MemoryMarshal.GetReference(pixelData), idx);
                 Vector256<uint> input = Unsafe.As<uint, Vector256<uint>>(ref pos);
                 Vector256<byte> a = Avx2.And(input.AsByte(), transformColorInverseAlphaGreenMask256);
-                Vector256<short> b = Avx2.ShuffleLow(a.AsInt16(), mmShuffle_2200);
-                Vector256<short> c = Avx2.ShuffleHigh(b.AsInt16(), mmShuffle_2200);
+                Vector256<short> b = Avx2.ShuffleLow(a.AsInt16(), SimdUtils.Shuffle.MMShuffle2200);
+                Vector256<short> c = Avx2.ShuffleHigh(b.AsInt16(), SimdUtils.Shuffle.MMShuffle2200);
                 Vector256<short> d = Avx2.MultiplyHigh(c.AsInt16(), multsrb.AsInt16());
                 Vector256<byte> e = Avx2.Add(input.AsByte(), d.AsByte());
                 Vector256<short> f = Avx2.ShiftLeftLogical(e.AsInt16(), 8);
@@ -496,20 +486,18 @@ internal static unsafe class LosslessUtils
         }
         else if (Sse2.IsSupported)
         {
-            var transformColorInverseAlphaGreenMask = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
+            Vector128<byte> transformColorInverseAlphaGreenMask = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255);
             Vector128<int> multsrb = MkCst16(Cst5b(m.GreenToRed), Cst5b(m.GreenToBlue));
             Vector128<int> multsb2 = MkCst16(Cst5b(m.RedToBlue), 0);
 
             nint idx;
             for (idx = 0; idx <= pixelData.Length - 4; idx += 4)
             {
-                const byte mmShuffle_2200 = 0b_10_10_00_00;
-
                 ref uint pos = ref Unsafe.Add(ref MemoryMarshal.GetReference(pixelData), idx);
                 Vector128<uint> input = Unsafe.As<uint, Vector128<uint>>(ref pos);
                 Vector128<byte> a = Sse2.And(input.AsByte(), transformColorInverseAlphaGreenMask);
-                Vector128<short> b = Sse2.ShuffleLow(a.AsInt16(), mmShuffle_2200);
-                Vector128<short> c = Sse2.ShuffleHigh(b.AsInt16(), mmShuffle_2200);
+                Vector128<short> b = Sse2.ShuffleLow(a.AsInt16(), SimdUtils.Shuffle.MMShuffle2200);
+                Vector128<short> c = Sse2.ShuffleHigh(b.AsInt16(), SimdUtils.Shuffle.MMShuffle2200);
                 Vector128<short> d = Sse2.MultiplyHigh(c.AsInt16(), multsrb.AsInt16());
                 Vector128<byte> e = Sse2.Add(input.AsByte(), d.AsByte());
                 Vector128<short> f = Sse2.ShiftLeftLogical(e.AsInt16(), 8);
diff --git a/src/ImageSharp/Formats/Webp/Lossy/Vp8Encoding.cs b/src/ImageSharp/Formats/Webp/Lossy/Vp8Encoding.cs
index cc263657f..82f00e876 100644
--- a/src/ImageSharp/Formats/Webp/Lossy/Vp8Encoding.cs
+++ b/src/ImageSharp/Formats/Webp/Lossy/Vp8Encoding.cs
@@ -521,9 +521,8 @@ internal static unsafe class Vp8Encoding
     {
         // *in01 = 00 01 10 11 02 03 12 13
         // *in23 = 20 21 30 31 22 23 32 33
-        const byte mmShuffle_2301 = 0b_10_11_00_01;
-        Vector128<short> shuf01_p = Sse2.ShuffleHigh(row01, mmShuffle_2301);
-        Vector128<short> shuf32_p = Sse2.ShuffleHigh(row23, mmShuffle_2301);
+        Vector128<short> shuf01_p = Sse2.ShuffleHigh(row01, SimdUtils.Shuffle.MMShuffle2301);
+        Vector128<short> shuf32_p = Sse2.ShuffleHigh(row23, SimdUtils.Shuffle.MMShuffle2301);
 
         // 00 01 10 11 03 02 13 12
         // 20 21 30 31 23 22 33 32
@@ -555,9 +554,7 @@ internal static unsafe class Vp8Encoding
         Vector128<short> shi = Sse2.UnpackHigh(s03, s12); // 2 3 2 3 2 3
         Vector128<int> v23 = Sse2.UnpackHigh(slo.AsInt32(), shi.AsInt32());
         out01 = Sse2.UnpackLow(slo.AsInt32(), shi.AsInt32());
-
-        const byte mmShuffle_1032 = 0b_01_00_11_10;
-        out32 = Sse2.Shuffle(v23, mmShuffle_1032);
+        out32 = Sse2.Shuffle(v23, SimdUtils.Shuffle.MMShuffle1032);
     }
 
     public static void FTransformPass2SSE2(Vector128<int> v01, Vector128<int> v32, Span<short> output)
diff --git a/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleFloat4Channel.cs b/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleFloat4Channel.cs
index bdeb88082..ec23364ca 100644
--- a/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleFloat4Channel.cs
+++ b/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleFloat4Channel.cs
@@ -9,7 +9,6 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk;
 [Config(typeof(Config.HwIntrinsics_SSE_AVX))]
 public class ShuffleFloat4Channel
 {
-    private static readonly byte Control = default(WXYZShuffle4).Control;
     private float[] source;
     private float[] destination;
 
@@ -25,9 +24,7 @@ public class ShuffleFloat4Channel
 
     [Benchmark]
     public void Shuffle4Channel()
-    {
-        SimdUtils.Shuffle4(this.source, this.destination, Control);
-    }
+        => SimdUtils.Shuffle4(this.source, this.destination, SimdUtils.Shuffle.MMShuffle2103);
 }
 
 // 2020-10-29
diff --git a/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs b/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs
index 9727731b6..2c4989f38 100644
--- a/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs
+++ b/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs
@@ -16,7 +16,7 @@ public partial class SimdUtilsTests
             // No need to test multiple shuffle controls as the
             // pipeline is always the same.
             int size = FeatureTestRunner.Deserialize<int>(serialized);
-            byte control = default(WZYXShuffle4).Control;
+            const byte control = SimdUtils.Shuffle.MMShuffle0123;
 
             TestShuffleFloat4Channel(
                 size,
@@ -45,39 +45,39 @@ public partial class SimdUtilsTests
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, wxyz),
-                wxyz.Control);
+                SimdUtils.Shuffle.MMShuffle2103);
 
             WZYXShuffle4 wzyx = default;
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, wzyx),
-                wzyx.Control);
+                SimdUtils.Shuffle.MMShuffle0123);
 
             YZWXShuffle4 yzwx = default;
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, yzwx),
-                yzwx.Control);
+                SimdUtils.Shuffle.MMShuffle0321);
 
             ZYXWShuffle4 zyxw = default;
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, zyxw),
-                zyxw.Control);
+                SimdUtils.Shuffle.MMShuffle3012);
 
-            var xwyz = new DefaultShuffle4(2, 1, 3, 0);
+            DefaultShuffle4 xwyz = new(2, 1, 3, 0);
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, xwyz),
                 xwyz.Control);
 
-            var yyyy = new DefaultShuffle4(1, 1, 1, 1);
+            DefaultShuffle4 yyyy = new(1, 1, 1, 1);
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, yyyy),
                 yyyy.Control);
 
-            var wwww = new DefaultShuffle4(3, 3, 3, 3);
+            DefaultShuffle4 wwww = new(3, 3, 3, 3);
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, wwww),
@@ -101,25 +101,25 @@ public partial class SimdUtilsTests
             // These cannot be expressed as a theory as you cannot
             // use RemoteExecutor within generic methods nor pass
             // IShuffle3 to the generic utils method.
-            var zyx = new DefaultShuffle3(0, 1, 2);
+            DefaultShuffle3 zyx = new(0, 1, 2);
             TestShuffleByte3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle3(s.Span, d.Span, zyx),
                 zyx.Control);
 
-            var xyz = new DefaultShuffle3(2, 1, 0);
+            DefaultShuffle3 xyz = new(2, 1, 0);
             TestShuffleByte3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle3(s.Span, d.Span, xyz),
                 xyz.Control);
 
-            var yyy = new DefaultShuffle3(1, 1, 1);
+            DefaultShuffle3 yyy = new(1, 1, 1);
             TestShuffleByte3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle3(s.Span, d.Span, yyy),
                 yyy.Control);
 
-            var zzz = new DefaultShuffle3(2, 2, 2);
+            DefaultShuffle3 zzz = new(2, 2, 2);
             TestShuffleByte3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle3(s.Span, d.Span, zzz),
@@ -147,21 +147,21 @@ public partial class SimdUtilsTests
             TestPad3Shuffle4Channel(
                 size,
                 (s, d) => SimdUtils.Pad3Shuffle4(s.Span, d.Span, xyzw),
-                xyzw.Control);
+                SimdUtils.Shuffle.MMShuffle3210);
 
-            var xwyz = new DefaultPad3Shuffle4(2, 1, 3, 0);
+            DefaultPad3Shuffle4 xwyz = new(2, 1, 3, 0);
             TestPad3Shuffle4Channel(
                 size,
                 (s, d) => SimdUtils.Pad3Shuffle4(s.Span, d.Span, xwyz),
                 xwyz.Control);
 
-            var yyyy = new DefaultPad3Shuffle4(1, 1, 1, 1);
+            DefaultPad3Shuffle4 yyyy = new(1, 1, 1, 1);
             TestPad3Shuffle4Channel(
                 size,
                 (s, d) => SimdUtils.Pad3Shuffle4(s.Span, d.Span, yyyy),
                 yyyy.Control);
 
-            var wwww = new DefaultPad3Shuffle4(3, 3, 3, 3);
+            DefaultPad3Shuffle4 wwww = new(3, 3, 3, 3);
             TestPad3Shuffle4Channel(
                 size,
                 (s, d) => SimdUtils.Pad3Shuffle4(s.Span, d.Span, wwww),
@@ -189,21 +189,21 @@ public partial class SimdUtilsTests
             TestShuffle4Slice3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4Slice3(s.Span, d.Span, xyzw),
-                xyzw.Control);
+                SimdUtils.Shuffle.MMShuffle3210);
 
-            var xwyz = new DefaultShuffle4Slice3(2, 1, 3, 0);
+            DefaultShuffle4Slice3 xwyz = new(2, 1, 3, 0);
             TestShuffle4Slice3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4Slice3(s.Span, d.Span, xwyz),
                 xwyz.Control);
 
-            var yyyy = new DefaultShuffle4Slice3(1, 1, 1, 1);
+            DefaultShuffle4Slice3 yyyy = new(1, 1, 1, 1);
             TestShuffle4Slice3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4Slice3(s.Span, d.Span, yyyy),
                 yyyy.Control);
 
-            var wwww = new DefaultShuffle4Slice3(3, 3, 3, 3);
+            DefaultShuffle4Slice3 wwww = new(3, 3, 3, 3);
             TestShuffle4Slice3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4Slice3(s.Span, d.Span, wwww),
@@ -222,7 +222,7 @@ public partial class SimdUtilsTests
         byte control)
     {
         float[] source = new Random(count).GenerateRandomFloatArray(count, 0, 256);
-        var result = new float[count];
+        float[] result = new float[count];
 
         float[] expected = new float[count];
 
@@ -253,7 +253,7 @@ public partial class SimdUtilsTests
     {
         byte[] source = new byte[count];
         new Random(count).NextBytes(source);
-        var result = new byte[count];
+        byte[] result = new byte[count];
 
         byte[] expected = new byte[count];
 
@@ -284,7 +284,7 @@ public partial class SimdUtilsTests
     {
         byte[] source = new byte[count];
         new Random(count).NextBytes(source);
-        var result = new byte[count];
+        byte[] result = new byte[count];
 
         byte[] expected = new byte[count];
 
@@ -315,7 +315,7 @@ public partial class SimdUtilsTests
         byte[] source = new byte[count];
         new Random(count).NextBytes(source);
 
-        var result = new byte[count * 4 / 3];
+        byte[] result = new byte[count * 4 / 3];
 
         byte[] expected = new byte[result.Length];
 
@@ -366,7 +366,7 @@ public partial class SimdUtilsTests
         byte[] source = new byte[count];
         new Random(count).NextBytes(source);
 
-        var result = new byte[count * 3 / 4];
+        byte[] result = new byte[count * 3 / 4];
 
         byte[] expected = new byte[result.Length];
 

From b9e4810418c47012aaed2dfa4c740b252c4c2585 Mon Sep 17 00:00:00 2001
From: James Jackson-South <james_south@hotmail.com>
Date: Tue, 21 Feb 2023 22:00:10 +1000
Subject: [PATCH 2/3] Refactor DefaultShuffleX to use constants.

---
 .../Helpers/Shuffle/IComponentShuffle.cs      | 24 +----
 .../Common/Helpers/Shuffle/IPad3Shuffle4.cs   | 24 +----
 .../Common/Helpers/Shuffle/IShuffle3.cs       | 20 +----
 .../Common/Helpers/Shuffle/IShuffle4Slice3.cs | 21 +----
 .../PixelFormats/Utils/PixelConverter.cs      | 88 ++++++++++++++++---
 .../Common/SimdUtilsTests.Shuffle.cs          | 26 +++---
 6 files changed, 103 insertions(+), 100 deletions(-)

diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs b/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs
index b93202814..6ce04ef21 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs
@@ -46,23 +46,10 @@ internal interface IShuffle4 : IComponentShuffle
 
 internal readonly struct DefaultShuffle4 : IShuffle4
 {
-    private readonly byte p3;
-    private readonly byte p2;
-    private readonly byte p1;
-    private readonly byte p0;
-
-    public DefaultShuffle4(byte p3, byte p2, byte p1, byte p0)
+    public DefaultShuffle4(byte control)
     {
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p3, 0, 3, nameof(p3));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p2, 0, 3, nameof(p2));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p1, 0, 3, nameof(p1));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p0, 0, 3, nameof(p0));
-
-        this.p3 = p3;
-        this.p2 = p2;
-        this.p1 = p1;
-        this.p0 = p0;
-        this.Control = Shuffle.MmShuffle(p3, p2, p1, p0);
+        DebugGuard.MustBeBetweenOrEqualTo<byte>(control, 0, 3, nameof(control));
+        this.Control = control;
     }
 
     public byte Control { get; }
@@ -77,10 +64,7 @@ internal readonly struct DefaultShuffle4 : IShuffle4
         ref byte sBase = ref MemoryMarshal.GetReference(source);
         ref byte dBase = ref MemoryMarshal.GetReference(dest);
 
-        int p3 = this.p3;
-        int p2 = this.p2;
-        int p1 = this.p1;
-        int p0 = this.p0;
+        Shuffle.InverseMmShuffle(this.Control, out int p3, out int p2, out int p1, out int p0);
 
         for (int i = 0; i < source.Length; i += 4)
         {
diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs b/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs
index 7f99b3a08..69a688e59 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs
@@ -14,23 +14,10 @@ internal interface IPad3Shuffle4 : IComponentShuffle
 
 internal readonly struct DefaultPad3Shuffle4 : IPad3Shuffle4
 {
-    private readonly byte p3;
-    private readonly byte p2;
-    private readonly byte p1;
-    private readonly byte p0;
-
-    public DefaultPad3Shuffle4(byte p3, byte p2, byte p1, byte p0)
+    public DefaultPad3Shuffle4(byte control)
     {
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p3, 0, 3, nameof(p3));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p2, 0, 3, nameof(p2));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p1, 0, 3, nameof(p1));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p0, 0, 3, nameof(p0));
-
-        this.p3 = p3;
-        this.p2 = p2;
-        this.p1 = p1;
-        this.p0 = p0;
-        this.Control = Shuffle.MmShuffle(p3, p2, p1, p0);
+        DebugGuard.MustBeBetweenOrEqualTo<byte>(control, 0, 3, nameof(control));
+        this.Control = control;
     }
 
     public byte Control { get; }
@@ -45,10 +32,7 @@ internal readonly struct DefaultPad3Shuffle4 : IPad3Shuffle4
         ref byte sBase = ref MemoryMarshal.GetReference(source);
         ref byte dBase = ref MemoryMarshal.GetReference(dest);
 
-        int p3 = this.p3;
-        int p2 = this.p2;
-        int p1 = this.p1;
-        int p0 = this.p0;
+        Shuffle.InverseMmShuffle(this.Control, out int p3, out int p2, out int p1, out int p0);
 
         Span<byte> temp = stackalloc byte[4];
         ref byte t = ref MemoryMarshal.GetReference(temp);
diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs
index edbd2cd5e..4e1c2a4e4 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs
@@ -14,20 +14,10 @@ internal interface IShuffle3 : IComponentShuffle
 
 internal readonly struct DefaultShuffle3 : IShuffle3
 {
-    private readonly byte p2;
-    private readonly byte p1;
-    private readonly byte p0;
-
-    public DefaultShuffle3(byte p2, byte p1, byte p0)
+    public DefaultShuffle3(byte control)
     {
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p2, 0, 2, nameof(p2));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p1, 0, 2, nameof(p1));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p0, 0, 2, nameof(p0));
-
-        this.p2 = p2;
-        this.p1 = p1;
-        this.p0 = p0;
-        this.Control = Shuffle.MmShuffle(3, p2, p1, p0);
+        DebugGuard.MustBeBetweenOrEqualTo<byte>(control, 0, 3, nameof(control));
+        this.Control = control;
     }
 
     public byte Control { get; }
@@ -42,9 +32,7 @@ internal readonly struct DefaultShuffle3 : IShuffle3
         ref byte sBase = ref MemoryMarshal.GetReference(source);
         ref byte dBase = ref MemoryMarshal.GetReference(dest);
 
-        int p2 = this.p2;
-        int p1 = this.p1;
-        int p0 = this.p0;
+        Shuffle.InverseMmShuffle(this.Control, out _, out int p2, out int p1, out int p0);
 
         for (int i = 0; i < source.Length; i += 3)
         {
diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs
index 2179a085b..ec0c4fcf2 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs
@@ -14,21 +14,10 @@ internal interface IShuffle4Slice3 : IComponentShuffle
 
 internal readonly struct DefaultShuffle4Slice3 : IShuffle4Slice3
 {
-    private readonly byte p2;
-    private readonly byte p1;
-    private readonly byte p0;
-
-    public DefaultShuffle4Slice3(byte p3, byte p2, byte p1, byte p0)
+    public DefaultShuffle4Slice3(byte control)
     {
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p3, 0, 3, nameof(p3));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p2, 0, 3, nameof(p2));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p1, 0, 3, nameof(p1));
-        DebugGuard.MustBeBetweenOrEqualTo<byte>(p0, 0, 3, nameof(p0));
-
-        this.p2 = p2;
-        this.p1 = p1;
-        this.p0 = p0;
-        this.Control = Shuffle.MmShuffle(p3, p2, p1, p0);
+        DebugGuard.MustBeBetweenOrEqualTo<byte>(control, 0, 3, nameof(control));
+        this.Control = control;
     }
 
     public byte Control { get; }
@@ -43,9 +32,7 @@ internal readonly struct DefaultShuffle4Slice3 : IShuffle4Slice3
         ref byte sBase = ref MemoryMarshal.GetReference(source);
         ref byte dBase = ref MemoryMarshal.GetReference(dest);
 
-        int p2 = this.p2;
-        int p1 = this.p1;
-        int p0 = this.p0;
+        Shuffle.InverseMmShuffle(this.Control, out _, out int p2, out int p1, out int p0);
 
         for (int i = 0, j = 0; i < dest.Length; i += 3, j += 4)
         {
diff --git a/src/ImageSharp/PixelFormats/Utils/PixelConverter.cs b/src/ImageSharp/PixelFormats/Utils/PixelConverter.cs
index 3ae5a3b5e..50a68906e 100644
--- a/src/ImageSharp/PixelFormats/Utils/PixelConverter.cs
+++ b/src/ImageSharp/PixelFormats/Utils/PixelConverter.cs
@@ -29,6 +29,8 @@ internal static class PixelConverter
         /// <see cref="Rgba32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Argb32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToArgb32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<WXYZShuffle4>(source, dest, default);
@@ -38,6 +40,8 @@ internal static class PixelConverter
         /// <see cref="Rgba32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgra32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgra32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<ZYXWShuffle4>(source, dest, default);
@@ -47,6 +51,8 @@ internal static class PixelConverter
         /// <see cref="Argb32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Abgr32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToAbgr32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<WZYXShuffle4>(source, dest, default);
@@ -56,6 +62,8 @@ internal static class PixelConverter
         /// <see cref="Rgba32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Rgb24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgb24(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4Slice3<XYZWShuffle4Slice3>(source, dest, default);
@@ -65,9 +73,11 @@ internal static class PixelConverter
         /// <see cref="Rgba32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgr24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgr24(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(3, 0, 1, 2));
+            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(SimdUtils.Shuffle.MMShuffle3012));
     }
 
     /// <summary>
@@ -82,6 +92,8 @@ internal static class PixelConverter
         /// <see cref="Argb32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Rgba32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgba32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<YZWXShuffle4>(source, dest, default);
@@ -91,6 +103,8 @@ internal static class PixelConverter
         /// <see cref="Argb32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgra32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgra32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<WZYXShuffle4>(source, dest, default);
@@ -100,6 +114,8 @@ internal static class PixelConverter
         /// <see cref="Argb32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Abgr32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToAbgr32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<XWZYShuffle4>(source, dest, default);
@@ -109,18 +125,22 @@ internal static class PixelConverter
         /// <see cref="Argb32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Rgb24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgb24(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(0, 3, 2, 1));
+            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(SimdUtils.Shuffle.MMShuffle0321));
 
         /// <summary>
         /// Converts a <see cref="ReadOnlySpan{Byte}"/> representing a collection of
         /// <see cref="Argb32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgr24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgr24(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(0, 1, 2, 3));
+            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(SimdUtils.Shuffle.MMShuffle0123));
     }
 
     /// <summary>
@@ -135,6 +155,8 @@ internal static class PixelConverter
         /// <see cref="Bgra32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Argb32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToArgb32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<WZYXShuffle4>(source, dest, default);
@@ -144,6 +166,8 @@ internal static class PixelConverter
         /// <see cref="Bgra32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Rgba32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgba32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<ZYXWShuffle4>(source, dest, default);
@@ -153,6 +177,8 @@ internal static class PixelConverter
         /// <see cref="Bgra32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Abgr32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToAbgr32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<WXYZShuffle4>(source, dest, default);
@@ -162,15 +188,19 @@ internal static class PixelConverter
         /// <see cref="Argb32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Rgb24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgb24(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(3, 0, 1, 2));
+            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(SimdUtils.Shuffle.MMShuffle3012));
 
         /// <summary>
         /// Converts a <see cref="ReadOnlySpan{Byte}"/> representing a collection of
         /// <see cref="Argb32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgr24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgr24(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4Slice3<XYZWShuffle4Slice3>(source, dest, default);
@@ -188,6 +218,8 @@ internal static class PixelConverter
         /// <see cref="Abgr32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Argb32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToArgb32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<XWZYShuffle4>(source, dest, default);
@@ -197,6 +229,8 @@ internal static class PixelConverter
         /// <see cref="Abgr32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgra32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgba32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<WZYXShuffle4>(source, dest, default);
@@ -206,6 +240,8 @@ internal static class PixelConverter
         /// <see cref="Abgr32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgra32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgra32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Shuffle4<YZWXShuffle4>(source, dest, default);
@@ -215,18 +251,22 @@ internal static class PixelConverter
         /// <see cref="Abgr32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Rgb24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgb24(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(0, 1, 2, 3));
+            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(SimdUtils.Shuffle.MMShuffle0123));
 
         /// <summary>
         /// Converts a <see cref="ReadOnlySpan{Byte}"/> representing a collection of
         /// <see cref="Abgr32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgr24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgr24(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(0, 3, 2, 1));
+            => SimdUtils.Shuffle4Slice3(source, dest, new DefaultShuffle4Slice3(SimdUtils.Shuffle.MMShuffle0321));
     }
 
     /// <summary>
@@ -241,6 +281,8 @@ internal static class PixelConverter
         /// <see cref="Rgb24"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Rgba32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgba32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Pad3Shuffle4<XYZWPad3Shuffle4>(source, dest, default);
@@ -250,36 +292,44 @@ internal static class PixelConverter
         /// <see cref="Rgba32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Argb32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToArgb32(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(2, 1, 0, 3));
+            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(SimdUtils.Shuffle.MMShuffle2103));
 
         /// <summary>
         /// Converts a <see cref="ReadOnlySpan{Byte}"/> representing a collection of
         /// <see cref="Rgba32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgra32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgra32(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(3, 0, 1, 2));
+            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(SimdUtils.Shuffle.MMShuffle3012));
 
         /// <summary>
         /// Converts a <see cref="ReadOnlySpan{Byte}"/> representing a collection of
         /// <see cref="Rgba32"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgra32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToAbgr32(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(0, 1, 2, 3));
+            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(SimdUtils.Shuffle.MMShuffle0123));
 
         /// <summary>
         /// Converts a <see cref="ReadOnlySpan{Byte}"/> representing a collection of
         /// <see cref="Rgb24"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgr24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgr24(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Shuffle3(source, dest, new DefaultShuffle3(0, 1, 2));
+            => SimdUtils.Shuffle3(source, dest, new DefaultShuffle3(SimdUtils.Shuffle.MMShuffle3012));
     }
 
     /// <summary>
@@ -294,24 +344,30 @@ internal static class PixelConverter
         /// <see cref="Bgr24"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Argb32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToArgb32(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(0, 1, 2, 3));
+            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(SimdUtils.Shuffle.MMShuffle0123));
 
         /// <summary>
         /// Converts a <see cref="ReadOnlySpan{Byte}"/> representing a collection of
         /// <see cref="Bgr24"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgra32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgba32(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(3, 0, 1, 2));
+            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(SimdUtils.Shuffle.MMShuffle3012));
 
         /// <summary>
         /// Converts a <see cref="ReadOnlySpan{Byte}"/> representing a collection of
         /// <see cref="Bgr24"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Bgra32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToBgra32(ReadOnlySpan<byte> source, Span<byte> dest)
             => SimdUtils.Pad3Shuffle4<XYZWPad3Shuffle4>(source, dest, default);
@@ -321,17 +377,21 @@ internal static class PixelConverter
         /// <see cref="Bgr24"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Abgr32"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToAbgr32(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(2, 1, 0, 3));
+            => SimdUtils.Pad3Shuffle4(source, dest, new DefaultPad3Shuffle4(SimdUtils.Shuffle.MMShuffle2103));
 
         /// <summary>
         /// Converts a <see cref="ReadOnlySpan{Byte}"/> representing a collection of
         /// <see cref="Bgr24"/> pixels to a <see cref="Span{Byte}"/> representing
         /// a collection of <see cref="Rgb24"/> pixels.
         /// </summary>
+        /// <param name="source">The source span of bytes.</param>
+        /// <param name="dest">The destination span of bytes.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static void ToRgb24(ReadOnlySpan<byte> source, Span<byte> dest)
-            => SimdUtils.Shuffle3(source, dest, new DefaultShuffle3(0, 1, 2));
+            => SimdUtils.Shuffle3(source, dest, new DefaultShuffle3(SimdUtils.Shuffle.MMShuffle3012));
     }
 }
diff --git a/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs b/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs
index 2c4989f38..5684c20e8 100644
--- a/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs
+++ b/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs
@@ -65,19 +65,19 @@ public partial class SimdUtilsTests
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, zyxw),
                 SimdUtils.Shuffle.MMShuffle3012);
 
-            DefaultShuffle4 xwyz = new(2, 1, 3, 0);
+            DefaultShuffle4 xwyz = new(SimdUtils.Shuffle.MMShuffle2130);
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, xwyz),
                 xwyz.Control);
 
-            DefaultShuffle4 yyyy = new(1, 1, 1, 1);
+            DefaultShuffle4 yyyy = new(SimdUtils.Shuffle.MMShuffle1111);
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, yyyy),
                 yyyy.Control);
 
-            DefaultShuffle4 wwww = new(3, 3, 3, 3);
+            DefaultShuffle4 wwww = new(SimdUtils.Shuffle.MMShuffle3333);
             TestShuffleByte4Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4(s.Span, d.Span, wwww),
@@ -101,25 +101,25 @@ public partial class SimdUtilsTests
             // These cannot be expressed as a theory as you cannot
             // use RemoteExecutor within generic methods nor pass
             // IShuffle3 to the generic utils method.
-            DefaultShuffle3 zyx = new(0, 1, 2);
+            DefaultShuffle3 zyx = new(SimdUtils.Shuffle.MMShuffle3012);
             TestShuffleByte3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle3(s.Span, d.Span, zyx),
                 zyx.Control);
 
-            DefaultShuffle3 xyz = new(2, 1, 0);
+            DefaultShuffle3 xyz = new(SimdUtils.Shuffle.MMShuffle3210);
             TestShuffleByte3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle3(s.Span, d.Span, xyz),
                 xyz.Control);
 
-            DefaultShuffle3 yyy = new(1, 1, 1);
+            DefaultShuffle3 yyy = new(SimdUtils.Shuffle.MMShuffle3111);
             TestShuffleByte3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle3(s.Span, d.Span, yyy),
                 yyy.Control);
 
-            DefaultShuffle3 zzz = new(2, 2, 2);
+            DefaultShuffle3 zzz = new(SimdUtils.Shuffle.MMShuffle3222);
             TestShuffleByte3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle3(s.Span, d.Span, zzz),
@@ -149,19 +149,19 @@ public partial class SimdUtilsTests
                 (s, d) => SimdUtils.Pad3Shuffle4(s.Span, d.Span, xyzw),
                 SimdUtils.Shuffle.MMShuffle3210);
 
-            DefaultPad3Shuffle4 xwyz = new(2, 1, 3, 0);
+            DefaultPad3Shuffle4 xwyz = new(SimdUtils.Shuffle.MMShuffle2130);
             TestPad3Shuffle4Channel(
                 size,
                 (s, d) => SimdUtils.Pad3Shuffle4(s.Span, d.Span, xwyz),
                 xwyz.Control);
 
-            DefaultPad3Shuffle4 yyyy = new(1, 1, 1, 1);
+            DefaultPad3Shuffle4 yyyy = new(SimdUtils.Shuffle.MMShuffle1111);
             TestPad3Shuffle4Channel(
                 size,
                 (s, d) => SimdUtils.Pad3Shuffle4(s.Span, d.Span, yyyy),
                 yyyy.Control);
 
-            DefaultPad3Shuffle4 wwww = new(3, 3, 3, 3);
+            DefaultPad3Shuffle4 wwww = new(SimdUtils.Shuffle.MMShuffle3333);
             TestPad3Shuffle4Channel(
                 size,
                 (s, d) => SimdUtils.Pad3Shuffle4(s.Span, d.Span, wwww),
@@ -191,19 +191,19 @@ public partial class SimdUtilsTests
                 (s, d) => SimdUtils.Shuffle4Slice3(s.Span, d.Span, xyzw),
                 SimdUtils.Shuffle.MMShuffle3210);
 
-            DefaultShuffle4Slice3 xwyz = new(2, 1, 3, 0);
+            DefaultShuffle4Slice3 xwyz = new(SimdUtils.Shuffle.MMShuffle2130);
             TestShuffle4Slice3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4Slice3(s.Span, d.Span, xwyz),
                 xwyz.Control);
 
-            DefaultShuffle4Slice3 yyyy = new(1, 1, 1, 1);
+            DefaultShuffle4Slice3 yyyy = new(SimdUtils.Shuffle.MMShuffle1111);
             TestShuffle4Slice3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4Slice3(s.Span, d.Span, yyyy),
                 yyyy.Control);
 
-            DefaultShuffle4Slice3 wwww = new(3, 3, 3, 3);
+            DefaultShuffle4Slice3 wwww = new(SimdUtils.Shuffle.MMShuffle3333);
             TestShuffle4Slice3Channel(
                 size,
                 (s, d) => SimdUtils.Shuffle4Slice3(s.Span, d.Span, wwww),

From eda0869f094c77cb61a8e4283c35f4ea3289bb47 Mon Sep 17 00:00:00 2001
From: James Jackson-South <james_south@hotmail.com>
Date: Wed, 22 Feb 2023 00:00:21 +1000
Subject: [PATCH 3/3] Update benchmarks and add more tests

---
 .../Helpers/Shuffle/IComponentShuffle.cs      |   2 +-
 .../Common/Helpers/Shuffle/IPad3Shuffle4.cs   |   2 +-
 .../Common/Helpers/Shuffle/IShuffle3.cs       |   2 +-
 .../Common/Helpers/Shuffle/IShuffle4Slice3.cs |   2 +-
 .../Common/Helpers/SimdUtils.HwIntrinsics.cs  |  10 +-
 .../Common/Helpers/SimdUtils.Shuffle.cs       |  10 +-
 .../Color/Bulk/Pad3Shuffle4Channel.cs         |  58 +++-
 .../Color/Bulk/Shuffle3Channel.cs             |  39 ++-
 .../Color/Bulk/Shuffle4Slice3Channel.cs       |  68 ++++-
 .../Color/Bulk/ShuffleByte4Channel.cs         |  39 ++-
 .../Color/Bulk/ShuffleFloat4Channel.cs        |  35 +++
 .../Common/SimdUtilsTests.Shuffle.cs          | 275 +++++++++++++++++-
 12 files changed, 500 insertions(+), 42 deletions(-)

diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs b/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs
index 6ce04ef21..345f4b5c2 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IComponentShuffle.cs
@@ -64,7 +64,7 @@ internal readonly struct DefaultShuffle4 : IShuffle4
         ref byte sBase = ref MemoryMarshal.GetReference(source);
         ref byte dBase = ref MemoryMarshal.GetReference(dest);
 
-        Shuffle.InverseMmShuffle(this.Control, out int p3, out int p2, out int p1, out int p0);
+        Shuffle.InverseMMShuffle(this.Control, out int p3, out int p2, out int p1, out int p0);
 
         for (int i = 0; i < source.Length; i += 4)
         {
diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs b/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs
index 69a688e59..348e8ec01 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IPad3Shuffle4.cs
@@ -32,7 +32,7 @@ internal readonly struct DefaultPad3Shuffle4 : IPad3Shuffle4
         ref byte sBase = ref MemoryMarshal.GetReference(source);
         ref byte dBase = ref MemoryMarshal.GetReference(dest);
 
-        Shuffle.InverseMmShuffle(this.Control, out int p3, out int p2, out int p1, out int p0);
+        Shuffle.InverseMMShuffle(this.Control, out int p3, out int p2, out int p1, out int p0);
 
         Span<byte> temp = stackalloc byte[4];
         ref byte t = ref MemoryMarshal.GetReference(temp);
diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs
index 4e1c2a4e4..6ac8a2428 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle3.cs
@@ -32,7 +32,7 @@ internal readonly struct DefaultShuffle3 : IShuffle3
         ref byte sBase = ref MemoryMarshal.GetReference(source);
         ref byte dBase = ref MemoryMarshal.GetReference(dest);
 
-        Shuffle.InverseMmShuffle(this.Control, out _, out int p2, out int p1, out int p0);
+        Shuffle.InverseMMShuffle(this.Control, out _, out int p2, out int p1, out int p0);
 
         for (int i = 0; i < source.Length; i += 3)
         {
diff --git a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs
index ec0c4fcf2..d4a5e8130 100644
--- a/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs
+++ b/src/ImageSharp/Common/Helpers/Shuffle/IShuffle4Slice3.cs
@@ -32,7 +32,7 @@ internal readonly struct DefaultShuffle4Slice3 : IShuffle4Slice3
         ref byte sBase = ref MemoryMarshal.GetReference(source);
         ref byte dBase = ref MemoryMarshal.GetReference(dest);
 
-        Shuffle.InverseMmShuffle(this.Control, out _, out int p2, out int p1, out int p0);
+        Shuffle.InverseMMShuffle(this.Control, out _, out int p2, out int p1, out int p0);
 
         for (int i = 0, j = 0; i < dest.Length; i += 3, j += 4)
         {
diff --git a/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs b/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs
index 7d2bab259..3841b64b4 100644
--- a/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs
+++ b/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs
@@ -297,7 +297,7 @@ internal static partial class SimdUtils
                 // shuffle controls to add to the library.
                 // We can add static ROS instances if need be in the future.
                 Span<byte> bytes = stackalloc byte[Vector256<byte>.Count];
-                Shuffle.MmShuffleSpan(ref bytes, control);
+                Shuffle.MMShuffleSpan(ref bytes, control);
                 Vector256<byte> vshuffle = Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(bytes));
 
                 ref Vector256<byte> sourceBase =
@@ -333,7 +333,7 @@ internal static partial class SimdUtils
             {
                 // Ssse3
                 Span<byte> bytes = stackalloc byte[Vector128<byte>.Count];
-                Shuffle.MmShuffleSpan(ref bytes, control);
+                Shuffle.MMShuffleSpan(ref bytes, control);
                 Vector128<byte> vshuffle = Unsafe.As<byte, Vector128<byte>>(ref MemoryMarshal.GetReference(bytes));
 
                 ref Vector128<byte> sourceBase =
@@ -382,7 +382,7 @@ internal static partial class SimdUtils
                 Vector128<byte> vmaske = Ssse3.AlignRight(vmasko, vmasko, 12);
 
                 Span<byte> bytes = stackalloc byte[Vector128<byte>.Count];
-                Shuffle.MmShuffleSpan(ref bytes, control);
+                Shuffle.MMShuffleSpan(ref bytes, control);
                 Vector128<byte> vshuffle = Unsafe.As<byte, Vector128<byte>>(ref MemoryMarshal.GetReference(bytes));
 
                 ref Vector128<byte> sourceBase =
@@ -445,7 +445,7 @@ internal static partial class SimdUtils
                 Vector128<byte> vfill = Vector128.Create(0xff000000ff000000ul).AsByte();
 
                 Span<byte> bytes = stackalloc byte[Vector128<byte>.Count];
-                Shuffle.MmShuffleSpan(ref bytes, control);
+                Shuffle.MMShuffleSpan(ref bytes, control);
                 Vector128<byte> vshuffle = Unsafe.As<byte, Vector128<byte>>(ref MemoryMarshal.GetReference(bytes));
 
                 ref Vector128<byte> sourceBase =
@@ -489,7 +489,7 @@ internal static partial class SimdUtils
                 Vector128<byte> vmaske = Ssse3.AlignRight(vmasko, vmasko, 12);
 
                 Span<byte> bytes = stackalloc byte[Vector128<byte>.Count];
-                Shuffle.MmShuffleSpan(ref bytes, control);
+                Shuffle.MMShuffleSpan(ref bytes, control);
                 Vector128<byte> vshuffle = Unsafe.As<byte, Vector128<byte>>(ref MemoryMarshal.GetReference(bytes));
 
                 ref Vector128<byte> sourceBase =
diff --git a/src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs b/src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs
index 4ff5f3d5b..aecdaa639 100644
--- a/src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs
+++ b/src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs
@@ -145,7 +145,7 @@ internal static partial class SimdUtils
     {
         ref float sBase = ref MemoryMarshal.GetReference(source);
         ref float dBase = ref MemoryMarshal.GetReference(dest);
-        Shuffle.InverseMmShuffle(control, out int p3, out int p2, out int p1, out int p0);
+        Shuffle.InverseMMShuffle(control, out int p3, out int p2, out int p1, out int p0);
 
         for (int i = 0; i < source.Length; i += 4)
         {
@@ -484,13 +484,13 @@ internal static partial class SimdUtils
         public const byte MMShuffle3333 = 0b11111111;
 
         [MethodImpl(InliningOptions.ShortMethod)]
-        public static byte MmShuffle(byte p3, byte p2, byte p1, byte p0)
+        public static byte MMShuffle(byte p3, byte p2, byte p1, byte p0)
             => (byte)((p3 << 6) | (p2 << 4) | (p1 << 2) | p0);
 
         [MethodImpl(InliningOptions.ShortMethod)]
-        public static void MmShuffleSpan(ref Span<byte> span, byte control)
+        public static void MMShuffleSpan(ref Span<byte> span, byte control)
         {
-            InverseMmShuffle(
+            InverseMMShuffle(
                  control,
                  out int p3,
                  out int p2,
@@ -509,7 +509,7 @@ internal static partial class SimdUtils
         }
 
         [MethodImpl(InliningOptions.ShortMethod)]
-        public static void InverseMmShuffle(
+        public static void InverseMMShuffle(
             byte control,
             out int p3,
             out int p2,
diff --git a/tests/ImageSharp.Benchmarks/Color/Bulk/Pad3Shuffle4Channel.cs b/tests/ImageSharp.Benchmarks/Color/Bulk/Pad3Shuffle4Channel.cs
index df308cdd4..73dcf55a1 100644
--- a/tests/ImageSharp.Benchmarks/Color/Bulk/Pad3Shuffle4Channel.cs
+++ b/tests/ImageSharp.Benchmarks/Color/Bulk/Pad3Shuffle4Channel.cs
@@ -8,8 +8,7 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk;
 [Config(typeof(Config.HwIntrinsics_SSE_AVX))]
 public class Pad3Shuffle4Channel
 {
-    private static readonly DefaultPad3Shuffle4 Control = new DefaultPad3Shuffle4(1, 0, 3, 2);
-    private static readonly XYZWPad3Shuffle4 ControlFast = default;
+    private static readonly DefaultPad3Shuffle4 Control = new(SimdUtils.Shuffle.MMShuffle1032);
     private byte[] source;
     private byte[] destination;
 
@@ -26,15 +25,11 @@ public class Pad3Shuffle4Channel
 
     [Benchmark]
     public void Pad3Shuffle4()
-    {
-        SimdUtils.Pad3Shuffle4(this.source, this.destination, Control);
-    }
+        => SimdUtils.Pad3Shuffle4(this.source, this.destination, Control);
 
     [Benchmark]
     public void Pad3Shuffle4FastFallback()
-    {
-        SimdUtils.Pad3Shuffle4(this.source, this.destination, ControlFast);
-    }
+        => SimdUtils.Pad3Shuffle4(this.source, this.destination, default(XYZWPad3Shuffle4));
 }
 
 // 2020-10-30
@@ -83,3 +78,50 @@ public class Pad3Shuffle4Channel
 // | Pad3Shuffle4FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1536 |   220.37 ns |  1.601 ns |  1.419 ns |   220.13 ns |  1.00 |    0.00 |     - |     - |     - |         - |
 // | Pad3Shuffle4FastFallback |             2. AVX |                                             Empty |  1536 |   111.54 ns |  2.173 ns |  2.901 ns |   111.27 ns |  0.51 |    0.01 |     - |     - |     - |         - |
 // | Pad3Shuffle4FastFallback |             3. SSE |                               COMPlus_EnableAVX=0 |  1536 |   110.23 ns |  0.456 ns |  0.427 ns |   110.25 ns |  0.50 |    0.00 |     - |     - |     - |         - |
+
+// 2023-02-21
+// ##########
+//
+// BenchmarkDotNet=v0.13.0, OS=Windows 10.0.22621
+// 11th Gen Intel Core i7-11370H 3.30GHz, 1 CPU, 8 logical and 4 physical cores
+// .NET SDK= 7.0.103
+//  [Host]             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  1. No HwIntrinsics : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  2. SSE             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  3. AVX             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+
+// Runtime=.NET 6.0
+
+// |                   Method |                Job |                              EnvironmentVariables | Count |      Mean |    Error |   StdDev | Ratio | Gen 0 | Gen 1 | Gen 2 | Allocated |
+// |------------------------- |------------------- |-------------------------------------------------- |------ |----------:|---------:|---------:|------:|------:|------:|------:|----------:|
+// |             Pad3Shuffle4 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |    96 |  57.45 ns | 0.126 ns | 0.118 ns |  1.00 |     - |     - |     - |         - |
+// |             Pad3Shuffle4 |             2. SSE |                               COMPlus_EnableAVX=0 |    96 |  14.70 ns | 0.105 ns | 0.098 ns |  0.26 |     - |     - |     - |         - |
+// |             Pad3Shuffle4 |             3. AVX |                                             Empty |    96 |  14.63 ns | 0.070 ns | 0.062 ns |  0.25 |     - |     - |     - |         - |
+// |                          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Pad3Shuffle4FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |    96 |  12.08 ns | 0.028 ns | 0.025 ns |  1.00 |     - |     - |     - |         - |
+// | Pad3Shuffle4FastFallback |             2. SSE |                               COMPlus_EnableAVX=0 |    96 |  14.04 ns | 0.050 ns | 0.044 ns |  1.16 |     - |     - |     - |         - |
+// | Pad3Shuffle4FastFallback |             3. AVX |                                             Empty |    96 |  13.90 ns | 0.086 ns | 0.080 ns |  1.15 |     - |     - |     - |         - |
+// |                          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// |             Pad3Shuffle4 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   384 | 202.67 ns | 2.010 ns | 1.678 ns |  1.00 |     - |     - |     - |         - |
+// |             Pad3Shuffle4 |             2. SSE |                               COMPlus_EnableAVX=0 |   384 |  25.54 ns | 0.060 ns | 0.053 ns |  0.13 |     - |     - |     - |         - |
+// |             Pad3Shuffle4 |             3. AVX |                                             Empty |   384 |  25.72 ns | 0.139 ns | 0.130 ns |  0.13 |     - |     - |     - |         - |
+// |                          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Pad3Shuffle4FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   384 |  60.35 ns | 0.080 ns | 0.071 ns |  1.00 |     - |     - |     - |         - |
+// | Pad3Shuffle4FastFallback |             2. SSE |                               COMPlus_EnableAVX=0 |   384 |  25.18 ns | 0.388 ns | 0.324 ns |  0.42 |     - |     - |     - |         - |
+// | Pad3Shuffle4FastFallback |             3. AVX |                                             Empty |   384 |  26.21 ns | 0.067 ns | 0.059 ns |  0.43 |     - |     - |     - |         - |
+// |                          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// |             Pad3Shuffle4 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   768 | 393.88 ns | 1.353 ns | 1.199 ns |  1.00 |     - |     - |     - |         - |
+// |             Pad3Shuffle4 |             2. SSE |                               COMPlus_EnableAVX=0 |   768 |  39.44 ns | 0.230 ns | 0.204 ns |  0.10 |     - |     - |     - |         - |
+// |             Pad3Shuffle4 |             3. AVX |                                             Empty |   768 |  39.51 ns | 0.108 ns | 0.101 ns |  0.10 |     - |     - |     - |         - |
+// |                          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Pad3Shuffle4FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   768 | 112.02 ns | 0.140 ns | 0.131 ns |  1.00 |     - |     - |     - |         - |
+// | Pad3Shuffle4FastFallback |             2. SSE |                               COMPlus_EnableAVX=0 |   768 |  38.60 ns | 0.091 ns | 0.080 ns |  0.34 |     - |     - |     - |         - |
+// | Pad3Shuffle4FastFallback |             3. AVX |                                             Empty |   768 |  38.18 ns | 0.100 ns | 0.084 ns |  0.34 |     - |     - |     - |         - |
+// |                          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// |             Pad3Shuffle4 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1536 | 777.95 ns | 1.719 ns | 1.342 ns |  1.00 |     - |     - |     - |         - |
+// |             Pad3Shuffle4 |             2. SSE |                               COMPlus_EnableAVX=0 |  1536 |  73.11 ns | 0.090 ns | 0.075 ns |  0.09 |     - |     - |     - |         - |
+// |             Pad3Shuffle4 |             3. AVX |                                             Empty |  1536 |  73.41 ns | 0.125 ns | 0.117 ns |  0.09 |     - |     - |     - |         - |
+// |                          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Pad3Shuffle4FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1536 | 218.14 ns | 0.377 ns | 0.334 ns |  1.00 |     - |     - |     - |         - |
+// | Pad3Shuffle4FastFallback |             2. SSE |                               COMPlus_EnableAVX=0 |  1536 |  72.55 ns | 1.418 ns | 1.184 ns |  0.33 |     - |     - |     - |         - |
+// | Pad3Shuffle4FastFallback |             3. AVX |                                             Empty |  1536 |  73.15 ns | 0.330 ns | 0.292 ns |  0.34 |     - |     - |     - |         - |
diff --git a/tests/ImageSharp.Benchmarks/Color/Bulk/Shuffle3Channel.cs b/tests/ImageSharp.Benchmarks/Color/Bulk/Shuffle3Channel.cs
index 7f6f5901f..9bd85f474 100644
--- a/tests/ImageSharp.Benchmarks/Color/Bulk/Shuffle3Channel.cs
+++ b/tests/ImageSharp.Benchmarks/Color/Bulk/Shuffle3Channel.cs
@@ -1,14 +1,16 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Runtime.InteropServices;
 using BenchmarkDotNet.Attributes;
+using Iced.Intel;
 
 namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk;
 
 [Config(typeof(Config.HwIntrinsics_SSE_AVX))]
 public class Shuffle3Channel
 {
-    private static readonly DefaultShuffle3 Control = new DefaultShuffle3(1, 0, 2);
+    private static readonly DefaultShuffle3 Control = new(SimdUtils.Shuffle.MMShuffle3102);
     private byte[] source;
     private byte[] destination;
 
@@ -25,9 +27,7 @@ public class Shuffle3Channel
 
     [Benchmark]
     public void Shuffle3()
-    {
-        SimdUtils.Shuffle3(this.source, this.destination, Control);
-    }
+        => SimdUtils.Shuffle3(this.source, this.destination, Control);
 }
 
 // 2020-11-02
@@ -60,3 +60,34 @@ public class Shuffle3Channel
 // |             Shuffle3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1536 | 773.70 ns | 5.516 ns | 4.890 ns | 772.96 ns |  1.00 |    0.00 |     - |     - |     - |         - |
 // |             Shuffle3 |             2. AVX |                                             Empty |  1536 | 190.41 ns | 1.090 ns | 0.851 ns | 190.38 ns |  0.25 |    0.00 |     - |     - |     - |         - |
 // |             Shuffle3 |             3. SSE |                               COMPlus_EnableAVX=0 |  1536 | 190.94 ns | 0.985 ns | 0.769 ns | 190.85 ns |  0.25 |    0.00 |     - |     - |     - |         - |
+
+// 2023-02-21
+// ##########
+//
+// BenchmarkDotNet=v0.13.0, OS=Windows 10.0.22621
+// 11th Gen Intel Core i7-11370H 3.30GHz, 1 CPU, 8 logical and 4 physical cores
+// .NET SDK= 7.0.103
+//  [Host]             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  1. No HwIntrinsics : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  2. SSE             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  3. AVX             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+
+// Runtime=.NET 6.0
+
+// |   Method |                Job |                              EnvironmentVariables | Count |      Mean |    Error |   StdDev | Ratio | Gen 0 | Gen 1 | Gen 2 | Allocated |
+// |--------- |------------------- |-------------------------------------------------- |------ |----------:|---------:|---------:|------:|------:|------:|------:|----------:|
+// | Shuffle3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |    96 |  44.55 ns | 0.564 ns | 0.528 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle3 |             2. SSE |                               COMPlus_EnableAVX=0 |    96 |  15.46 ns | 0.064 ns | 0.060 ns |  0.35 |     - |     - |     - |         - |
+// | Shuffle3 |             3. AVX |                                             Empty |    96 |  15.18 ns | 0.056 ns | 0.053 ns |  0.34 |     - |     - |     - |         - |
+// |          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Shuffle3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   384 | 155.68 ns | 0.539 ns | 0.504 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle3 |             2. SSE |                               COMPlus_EnableAVX=0 |   384 |  30.04 ns | 0.100 ns | 0.089 ns |  0.19 |     - |     - |     - |         - |
+// | Shuffle3 |             3. AVX |                                             Empty |   384 |  29.70 ns | 0.061 ns | 0.054 ns |  0.19 |     - |     - |     - |         - |
+// |          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Shuffle3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   768 | 302.76 ns | 1.023 ns | 0.957 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle3 |             2. SSE |                               COMPlus_EnableAVX=0 |   768 |  50.24 ns | 0.098 ns | 0.092 ns |  0.17 |     - |     - |     - |         - |
+// | Shuffle3 |             3. AVX |                                             Empty |   768 |  49.28 ns | 0.156 ns | 0.131 ns |  0.16 |     - |     - |     - |         - |
+// |          |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Shuffle3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1536 | 596.53 ns | 2.675 ns | 2.503 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle3 |             2. SSE |                               COMPlus_EnableAVX=0 |  1536 |  94.09 ns | 0.312 ns | 0.260 ns |  0.16 |     - |     - |     - |         - |
+// | Shuffle3 |             3. AVX |                                             Empty |  1536 |  93.57 ns | 0.196 ns | 0.183 ns |  0.16 |     - |     - |     - |         - |
diff --git a/tests/ImageSharp.Benchmarks/Color/Bulk/Shuffle4Slice3Channel.cs b/tests/ImageSharp.Benchmarks/Color/Bulk/Shuffle4Slice3Channel.cs
index 0a0afb705..f0890221c 100644
--- a/tests/ImageSharp.Benchmarks/Color/Bulk/Shuffle4Slice3Channel.cs
+++ b/tests/ImageSharp.Benchmarks/Color/Bulk/Shuffle4Slice3Channel.cs
@@ -1,15 +1,16 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Runtime.InteropServices;
 using BenchmarkDotNet.Attributes;
+using Iced.Intel;
 
 namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk;
 
 [Config(typeof(Config.HwIntrinsics_SSE_AVX))]
 public class Shuffle4Slice3Channel
 {
-    private static readonly DefaultShuffle4Slice3 Control = new DefaultShuffle4Slice3(1, 0, 3, 2);
-    private static readonly XYZWShuffle4Slice3 ControlFast = default;
+    private static readonly DefaultShuffle4Slice3 Control = new(SimdUtils.Shuffle.MMShuffle1032);
     private byte[] source;
     private byte[] destination;
 
@@ -26,15 +27,11 @@ public class Shuffle4Slice3Channel
 
     [Benchmark]
     public void Shuffle4Slice3()
-    {
-        SimdUtils.Shuffle4Slice3(this.source, this.destination, Control);
-    }
+        => SimdUtils.Shuffle4Slice3(this.source, this.destination, Control);
 
     [Benchmark]
     public void Shuffle4Slice3FastFallback()
-    {
-        SimdUtils.Shuffle4Slice3(this.source, this.destination, ControlFast);
-    }
+        => SimdUtils.Shuffle4Slice3(this.source, this.destination, default(XYZWShuffle4Slice3));
 }
 
 // 2020-10-29
@@ -91,3 +88,58 @@ public class Shuffle4Slice3Channel
 // | Shuffle4Slice3FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  2048 | 382.97 ns | 1.064 ns | 0.831 ns | 382.87 ns |  1.00 |    0.00 |     - |     - |     - |         - |
 // | Shuffle4Slice3FastFallback |             2. AVX |                                             Empty |  2048 | 126.93 ns | 0.382 ns | 0.339 ns | 126.94 ns |  0.33 |    0.00 |     - |     - |     - |         - |
 // | Shuffle4Slice3FastFallback |             3. SSE |                               COMPlus_EnableAVX=0 |  2048 | 149.36 ns | 1.875 ns | 1.754 ns | 149.33 ns |  0.39 |    0.00 |     - |     - |     - |         - |
+
+// 2023-02-21
+// ##########
+//
+// BenchmarkDotNet=v0.13.0, OS=Windows 10.0.22621
+// 11th Gen Intel Core i7-11370H 3.30GHz, 1 CPU, 8 logical and 4 physical cores
+// .NET SDK= 7.0.103
+//  [Host]             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  1. No HwIntrinsics : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  2. SSE             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  3. AVX             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//
+// Runtime=.NET 6.0
+//
+// |                     Method |                Job |                              EnvironmentVariables | Count |      Mean |    Error |   StdDev | Ratio | Gen 0 | Gen 1 | Gen 2 | Allocated |
+// |--------------------------- |------------------- |-------------------------------------------------- |------ |----------:|---------:|---------:|------:|------:|------:|------:|----------:|
+// |             Shuffle4Slice3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   128 |  45.59 ns | 0.166 ns | 0.147 ns |  1.00 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             2. SSE |                               COMPlus_EnableAVX=0 |   128 |  15.62 ns | 0.056 ns | 0.052 ns |  0.34 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             3. AVX |                                             Empty |   128 |  16.37 ns | 0.047 ns | 0.040 ns |  0.36 |     - |     - |     - |         - |
+// |                            |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Shuffle4Slice3FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   128 |  13.23 ns | 0.028 ns | 0.026 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             2. SSE |                               COMPlus_EnableAVX=0 |   128 |  14.41 ns | 0.013 ns | 0.012 ns |  1.09 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             3. AVX |                                             Empty |   128 |  14.70 ns | 0.050 ns | 0.047 ns |  1.11 |     - |     - |     - |         - |
+// |                            |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// |             Shuffle4Slice3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   256 |  85.48 ns | 0.192 ns | 0.179 ns |  1.00 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             2. SSE |                               COMPlus_EnableAVX=0 |   256 |  19.18 ns | 0.230 ns | 0.204 ns |  0.22 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             3. AVX |                                             Empty |   256 |  18.66 ns | 0.017 ns | 0.015 ns |  0.22 |     - |     - |     - |         - |
+// |                            |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Shuffle4Slice3FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   256 |  24.34 ns | 0.078 ns | 0.073 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             2. SSE |                               COMPlus_EnableAVX=0 |   256 |  18.58 ns | 0.061 ns | 0.057 ns |  0.76 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             3. AVX |                                             Empty |   256 |  19.23 ns | 0.018 ns | 0.016 ns |  0.79 |     - |     - |     - |         - |
+// |                            |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// |             Shuffle4Slice3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   512 | 165.31 ns | 0.742 ns | 0.694 ns |  1.00 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             2. SSE |                               COMPlus_EnableAVX=0 |   512 |  28.10 ns | 0.077 ns | 0.068 ns |  0.17 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             3. AVX |                                             Empty |   512 |  28.99 ns | 0.018 ns | 0.014 ns |  0.18 |     - |     - |     - |         - |
+// |                            |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Shuffle4Slice3FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   512 |  53.45 ns | 0.270 ns | 0.226 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             2. SSE |                               COMPlus_EnableAVX=0 |   512 |  27.50 ns | 0.034 ns | 0.028 ns |  0.51 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             3. AVX |                                             Empty |   512 |  28.76 ns | 0.017 ns | 0.015 ns |  0.54 |     - |     - |     - |         - |
+// |                            |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// |             Shuffle4Slice3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1024 | 323.87 ns | 0.549 ns | 0.487 ns |  1.00 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             2. SSE |                               COMPlus_EnableAVX=0 |  1024 |  40.81 ns | 0.056 ns | 0.050 ns |  0.13 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             3. AVX |                                             Empty |  1024 |  39.95 ns | 0.075 ns | 0.067 ns |  0.12 |     - |     - |     - |         - |
+// |                            |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Shuffle4Slice3FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1024 | 101.37 ns | 0.080 ns | 0.067 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             2. SSE |                               COMPlus_EnableAVX=0 |  1024 |  40.72 ns | 0.049 ns | 0.041 ns |  0.40 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             3. AVX |                                             Empty |  1024 |  39.78 ns | 0.029 ns | 0.027 ns |  0.39 |     - |     - |     - |         - |
+// |                            |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// |             Shuffle4Slice3 | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  2048 | 642.95 ns | 2.067 ns | 1.933 ns |  1.00 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             2. SSE |                               COMPlus_EnableAVX=0 |  2048 |  73.19 ns | 0.082 ns | 0.077 ns |  0.11 |     - |     - |     - |         - |
+// |             Shuffle4Slice3 |             3. AVX |                                             Empty |  2048 |  69.83 ns | 0.319 ns | 0.267 ns |  0.11 |     - |     - |     - |         - |
+// |                            |                    |                                                   |       |           |          |          |       |       |       |       |           |
+// | Shuffle4Slice3FastFallback | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  2048 | 196.85 ns | 0.238 ns | 0.211 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             2. SSE |                               COMPlus_EnableAVX=0 |  2048 |  72.89 ns | 0.117 ns | 0.098 ns |  0.37 |     - |     - |     - |         - |
+// | Shuffle4Slice3FastFallback |             3. AVX |                                             Empty |  2048 |  69.59 ns | 0.073 ns | 0.061 ns |  0.35 |     - |     - |     - |         - |
diff --git a/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleByte4Channel.cs b/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleByte4Channel.cs
index 6323e1c55..8e740743b 100644
--- a/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleByte4Channel.cs
+++ b/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleByte4Channel.cs
@@ -24,9 +24,7 @@ public class ShuffleByte4Channel
 
     [Benchmark]
     public void Shuffle4Channel()
-    {
-        SimdUtils.Shuffle4<WXYZShuffle4>(this.source, this.destination, default);
-    }
+        => SimdUtils.Shuffle4<WXYZShuffle4>(this.source, this.destination, default);
 }
 
 // 2020-10-29
@@ -63,3 +61,38 @@ public class ShuffleByte4Channel
 // | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  2048 | 315.29 ns | 5.206 ns | 6.583 ns |  1.00 |    0.00 |     - |     - |     - |         - |
 // | Shuffle4Channel |             2. AVX |                                             Empty |  2048 |  57.37 ns | 1.152 ns | 1.078 ns |  0.18 |    0.01 |     - |     - |     - |         - |
 // | Shuffle4Channel |             3. SSE |                               COMPlus_EnableAVX=0 |  2048 |  65.75 ns | 1.198 ns | 1.600 ns |  0.21 |    0.01 |     - |     - |     - |         - |
+
+// 2023-02-21
+// ##########
+//
+// BenchmarkDotNet=v0.13.0, OS=Windows 10.0.22621
+// 11th Gen Intel Core i7-11370H 3.30GHz, 1 CPU, 8 logical and 4 physical cores
+// .NET SDK= 7.0.103
+//  [Host]             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  1. No HwIntrinsics : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  2. SSE             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  3. AVX             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//
+// Runtime=.NET 6.0
+//
+// |          Method |                Job |                              EnvironmentVariables | Count |      Mean |    Error |   StdDev | Ratio | RatioSD | Gen 0 | Gen 1 | Gen 2 | Allocated |
+// |---------------- |------------------- |-------------------------------------------------- |------ |----------:|---------:|---------:|------:|--------:|------:|------:|------:|----------:|
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   128 |  10.76 ns | 0.033 ns | 0.029 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |   128 |  11.39 ns | 0.045 ns | 0.040 ns |  1.06 |    0.01 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |   128 |  14.05 ns | 0.029 ns | 0.024 ns |  1.31 |    0.00 |     - |     - |     - |         - |
+// |                 |                    |                                                   |       |           |          |          |       |         |       |       |       |           |
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   256 |  32.09 ns | 0.655 ns | 1.000 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |   256 |  14.03 ns | 0.047 ns | 0.041 ns |  0.44 |    0.02 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |   256 |  15.18 ns | 0.052 ns | 0.043 ns |  0.48 |    0.03 |     - |     - |     - |         - |
+// |                 |                    |                                                   |       |           |          |          |       |         |       |       |       |           |
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   512 |  59.26 ns | 0.084 ns | 0.070 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |   512 |  18.80 ns | 0.036 ns | 0.034 ns |  0.32 |    0.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |   512 |  17.69 ns | 0.038 ns | 0.034 ns |  0.30 |    0.00 |     - |     - |     - |         - |
+// |                 |                    |                                                   |       |           |          |          |       |         |       |       |       |           |
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1024 | 112.48 ns | 0.285 ns | 0.253 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |  1024 |  31.57 ns | 0.041 ns | 0.036 ns |  0.28 |    0.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |  1024 |  28.41 ns | 0.068 ns | 0.064 ns |  0.25 |    0.00 |     - |     - |     - |         - |
+// |                 |                    |                                                   |       |           |          |          |       |         |       |       |       |           |
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  2048 | 218.59 ns | 0.303 ns | 0.283 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |  2048 |  53.04 ns | 0.106 ns | 0.099 ns |  0.24 |    0.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |  2048 |  34.74 ns | 0.061 ns | 0.054 ns |  0.16 |    0.00 |     - |     - |     - |         - |
diff --git a/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleFloat4Channel.cs b/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleFloat4Channel.cs
index ec23364ca..1c338e1a6 100644
--- a/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleFloat4Channel.cs
+++ b/tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleFloat4Channel.cs
@@ -61,3 +61,38 @@ public class ShuffleFloat4Channel
 // | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  2048 | 980.134 ns | 3.7407 ns | 3.1237 ns |  1.00 |     - |     - |     - |         - |
 // | Shuffle4Channel |             2. AVX |                                             Empty |  2048 | 105.120 ns | 0.6140 ns | 0.5443 ns |  0.11 |     - |     - |     - |         - |
 // | Shuffle4Channel |             3. SSE |                               COMPlus_EnableAVX=0 |  2048 | 216.473 ns | 2.3268 ns | 2.0627 ns |  0.22 |     - |     - |     - |         - |
+
+// 2023-02-21
+// ##########
+//
+// BenchmarkDotNet=v0.13.0, OS=Windows 10.0.22621
+// 11th Gen Intel Core i7-11370H 3.30GHz, 1 CPU, 8 logical and 4 physical cores
+// .NET SDK= 7.0.103
+//  [Host]             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  1. No HwIntrinsics : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  2. SSE             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//  3. AVX             : .NET 6.0.14 (6.0.1423.7309), X64 RyuJIT
+//
+// Runtime=.NET 6.0
+//
+// |          Method |                Job |                              EnvironmentVariables | Count |       Mean |     Error |    StdDev | Ratio | Gen 0 | Gen 1 | Gen 2 | Allocated |
+// |---------------- |------------------- |-------------------------------------------------- |------ |-----------:|----------:|----------:|------:|------:|------:|------:|----------:|
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   128 |  57.819 ns | 0.2360 ns | 0.1970 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |   128 |  11.564 ns | 0.0234 ns | 0.0195 ns |  0.20 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |   128 |   7.770 ns | 0.0696 ns | 0.0617 ns |  0.13 |     - |     - |     - |         - |
+// |                 |                    |                                                   |       |            |           |           |       |       |       |       |           |
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   256 | 105.282 ns | 0.2713 ns | 0.2405 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |   256 |  19.867 ns | 0.0393 ns | 0.0348 ns |  0.19 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |   256 |  17.586 ns | 0.0582 ns | 0.0544 ns |  0.17 |     - |     - |     - |         - |
+// |                 |                    |                                                   |       |            |           |           |       |       |       |       |           |
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   512 | 200.799 ns | 0.5678 ns | 0.5033 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |   512 |  41.137 ns | 0.1524 ns | 0.1351 ns |  0.20 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |   512 |  24.040 ns | 0.0445 ns | 0.0395 ns |  0.12 |     - |     - |     - |         - |
+// |                 |                    |                                                   |       |            |           |           |       |       |       |       |           |
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1024 | 401.046 ns | 0.5865 ns | 0.5199 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |  1024 |  94.904 ns | 0.4633 ns | 0.4334 ns |  0.24 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |  1024 |  68.456 ns | 0.1192 ns | 0.0996 ns |  0.17 |     - |     - |     - |         - |
+// |                 |                    |                                                   |       |            |           |           |       |       |       |       |           |
+// | Shuffle4Channel | 1. No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  2048 | 772.297 ns | 0.6270 ns | 0.5558 ns |  1.00 |     - |     - |     - |         - |
+// | Shuffle4Channel |             2. SSE |                               COMPlus_EnableAVX=0 |  2048 | 184.561 ns | 0.4319 ns | 0.4040 ns |  0.24 |     - |     - |     - |         - |
+// | Shuffle4Channel |             3. AVX |                                             Empty |  2048 | 133.634 ns | 1.7864 ns | 1.8345 ns |  0.17 |     - |     - |     - |         - |
diff --git a/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs b/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs
index 5684c20e8..960b0613e 100644
--- a/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs
+++ b/tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs
@@ -7,6 +7,271 @@ namespace SixLabors.ImageSharp.Tests.Common;
 
 public partial class SimdUtilsTests
 {
+    public static readonly TheoryData<byte, byte> MMShuffleData = new()
+    {
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 0, 0), SimdUtils.Shuffle.MMShuffle0000 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 0, 1), SimdUtils.Shuffle.MMShuffle0001 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 0, 2), SimdUtils.Shuffle.MMShuffle0002 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 0, 3), SimdUtils.Shuffle.MMShuffle0003 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 1, 0), SimdUtils.Shuffle.MMShuffle0010 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 1, 1), SimdUtils.Shuffle.MMShuffle0011 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 1, 2), SimdUtils.Shuffle.MMShuffle0012 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 1, 3), SimdUtils.Shuffle.MMShuffle0013 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 2, 0), SimdUtils.Shuffle.MMShuffle0020 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 2, 1), SimdUtils.Shuffle.MMShuffle0021 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 2, 2), SimdUtils.Shuffle.MMShuffle0022 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 2, 3), SimdUtils.Shuffle.MMShuffle0023 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 3, 0), SimdUtils.Shuffle.MMShuffle0030 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 3, 1), SimdUtils.Shuffle.MMShuffle0031 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 3, 2), SimdUtils.Shuffle.MMShuffle0032 },
+        { SimdUtils.Shuffle.MMShuffle(0, 0, 3, 3), SimdUtils.Shuffle.MMShuffle0033 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 0, 0), SimdUtils.Shuffle.MMShuffle0100 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 0, 1), SimdUtils.Shuffle.MMShuffle0101 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 0, 2), SimdUtils.Shuffle.MMShuffle0102 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 0, 3), SimdUtils.Shuffle.MMShuffle0103 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 1, 0), SimdUtils.Shuffle.MMShuffle0110 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 1, 1), SimdUtils.Shuffle.MMShuffle0111 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 1, 2), SimdUtils.Shuffle.MMShuffle0112 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 1, 3), SimdUtils.Shuffle.MMShuffle0113 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 2, 0), SimdUtils.Shuffle.MMShuffle0120 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 2, 1), SimdUtils.Shuffle.MMShuffle0121 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 2, 2), SimdUtils.Shuffle.MMShuffle0122 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 2, 3), SimdUtils.Shuffle.MMShuffle0123 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 3, 0), SimdUtils.Shuffle.MMShuffle0130 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 3, 1), SimdUtils.Shuffle.MMShuffle0131 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 3, 2), SimdUtils.Shuffle.MMShuffle0132 },
+        { SimdUtils.Shuffle.MMShuffle(0, 1, 3, 3), SimdUtils.Shuffle.MMShuffle0133 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 0, 0), SimdUtils.Shuffle.MMShuffle0200 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 0, 1), SimdUtils.Shuffle.MMShuffle0201 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 0, 2), SimdUtils.Shuffle.MMShuffle0202 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 0, 3), SimdUtils.Shuffle.MMShuffle0203 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 1, 0), SimdUtils.Shuffle.MMShuffle0210 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 1, 1), SimdUtils.Shuffle.MMShuffle0211 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 1, 2), SimdUtils.Shuffle.MMShuffle0212 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 1, 3), SimdUtils.Shuffle.MMShuffle0213 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 2, 0), SimdUtils.Shuffle.MMShuffle0220 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 2, 1), SimdUtils.Shuffle.MMShuffle0221 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 2, 2), SimdUtils.Shuffle.MMShuffle0222 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 2, 3), SimdUtils.Shuffle.MMShuffle0223 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 3, 0), SimdUtils.Shuffle.MMShuffle0230 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 3, 1), SimdUtils.Shuffle.MMShuffle0231 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 3, 2), SimdUtils.Shuffle.MMShuffle0232 },
+        { SimdUtils.Shuffle.MMShuffle(0, 2, 3, 3), SimdUtils.Shuffle.MMShuffle0233 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 0, 0), SimdUtils.Shuffle.MMShuffle0300 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 0, 1), SimdUtils.Shuffle.MMShuffle0301 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 0, 2), SimdUtils.Shuffle.MMShuffle0302 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 0, 3), SimdUtils.Shuffle.MMShuffle0303 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 1, 0), SimdUtils.Shuffle.MMShuffle0310 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 1, 1), SimdUtils.Shuffle.MMShuffle0311 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 1, 2), SimdUtils.Shuffle.MMShuffle0312 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 1, 3), SimdUtils.Shuffle.MMShuffle0313 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 2, 0), SimdUtils.Shuffle.MMShuffle0320 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 2, 1), SimdUtils.Shuffle.MMShuffle0321 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 2, 2), SimdUtils.Shuffle.MMShuffle0322 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 2, 3), SimdUtils.Shuffle.MMShuffle0323 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 3, 0), SimdUtils.Shuffle.MMShuffle0330 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 3, 1), SimdUtils.Shuffle.MMShuffle0331 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 3, 2), SimdUtils.Shuffle.MMShuffle0332 },
+        { SimdUtils.Shuffle.MMShuffle(0, 3, 3, 3), SimdUtils.Shuffle.MMShuffle0333 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 0, 0), SimdUtils.Shuffle.MMShuffle1000 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 0, 1), SimdUtils.Shuffle.MMShuffle1001 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 0, 2), SimdUtils.Shuffle.MMShuffle1002 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 0, 3), SimdUtils.Shuffle.MMShuffle1003 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 1, 0), SimdUtils.Shuffle.MMShuffle1010 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 1, 1), SimdUtils.Shuffle.MMShuffle1011 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 1, 2), SimdUtils.Shuffle.MMShuffle1012 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 1, 3), SimdUtils.Shuffle.MMShuffle1013 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 2, 0), SimdUtils.Shuffle.MMShuffle1020 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 2, 1), SimdUtils.Shuffle.MMShuffle1021 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 2, 2), SimdUtils.Shuffle.MMShuffle1022 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 2, 3), SimdUtils.Shuffle.MMShuffle1023 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 3, 0), SimdUtils.Shuffle.MMShuffle1030 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 3, 1), SimdUtils.Shuffle.MMShuffle1031 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 3, 2), SimdUtils.Shuffle.MMShuffle1032 },
+        { SimdUtils.Shuffle.MMShuffle(1, 0, 3, 3), SimdUtils.Shuffle.MMShuffle1033 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 0, 0), SimdUtils.Shuffle.MMShuffle1100 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 0, 1), SimdUtils.Shuffle.MMShuffle1101 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 0, 2), SimdUtils.Shuffle.MMShuffle1102 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 0, 3), SimdUtils.Shuffle.MMShuffle1103 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 1, 0), SimdUtils.Shuffle.MMShuffle1110 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 1, 1), SimdUtils.Shuffle.MMShuffle1111 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 1, 2), SimdUtils.Shuffle.MMShuffle1112 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 1, 3), SimdUtils.Shuffle.MMShuffle1113 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 2, 0), SimdUtils.Shuffle.MMShuffle1120 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 2, 1), SimdUtils.Shuffle.MMShuffle1121 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 2, 2), SimdUtils.Shuffle.MMShuffle1122 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 2, 3), SimdUtils.Shuffle.MMShuffle1123 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 3, 0), SimdUtils.Shuffle.MMShuffle1130 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 3, 1), SimdUtils.Shuffle.MMShuffle1131 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 3, 2), SimdUtils.Shuffle.MMShuffle1132 },
+        { SimdUtils.Shuffle.MMShuffle(1, 1, 3, 3), SimdUtils.Shuffle.MMShuffle1133 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 0, 0), SimdUtils.Shuffle.MMShuffle1200 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 0, 1), SimdUtils.Shuffle.MMShuffle1201 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 0, 2), SimdUtils.Shuffle.MMShuffle1202 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 0, 3), SimdUtils.Shuffle.MMShuffle1203 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 1, 0), SimdUtils.Shuffle.MMShuffle1210 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 1, 1), SimdUtils.Shuffle.MMShuffle1211 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 1, 2), SimdUtils.Shuffle.MMShuffle1212 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 1, 3), SimdUtils.Shuffle.MMShuffle1213 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 2, 0), SimdUtils.Shuffle.MMShuffle1220 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 2, 1), SimdUtils.Shuffle.MMShuffle1221 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 2, 2), SimdUtils.Shuffle.MMShuffle1222 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 2, 3), SimdUtils.Shuffle.MMShuffle1223 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 3, 0), SimdUtils.Shuffle.MMShuffle1230 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 3, 1), SimdUtils.Shuffle.MMShuffle1231 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 3, 2), SimdUtils.Shuffle.MMShuffle1232 },
+        { SimdUtils.Shuffle.MMShuffle(1, 2, 3, 3), SimdUtils.Shuffle.MMShuffle1233 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 0, 0), SimdUtils.Shuffle.MMShuffle1300 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 0, 1), SimdUtils.Shuffle.MMShuffle1301 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 0, 2), SimdUtils.Shuffle.MMShuffle1302 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 0, 3), SimdUtils.Shuffle.MMShuffle1303 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 1, 0), SimdUtils.Shuffle.MMShuffle1310 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 1, 1), SimdUtils.Shuffle.MMShuffle1311 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 1, 2), SimdUtils.Shuffle.MMShuffle1312 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 1, 3), SimdUtils.Shuffle.MMShuffle1313 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 2, 0), SimdUtils.Shuffle.MMShuffle1320 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 2, 1), SimdUtils.Shuffle.MMShuffle1321 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 2, 2), SimdUtils.Shuffle.MMShuffle1322 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 2, 3), SimdUtils.Shuffle.MMShuffle1323 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 3, 0), SimdUtils.Shuffle.MMShuffle1330 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 3, 1), SimdUtils.Shuffle.MMShuffle1331 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 3, 2), SimdUtils.Shuffle.MMShuffle1332 },
+        { SimdUtils.Shuffle.MMShuffle(1, 3, 3, 3), SimdUtils.Shuffle.MMShuffle1333 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 0, 0), SimdUtils.Shuffle.MMShuffle2000 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 0, 1), SimdUtils.Shuffle.MMShuffle2001 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 0, 2), SimdUtils.Shuffle.MMShuffle2002 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 0, 3), SimdUtils.Shuffle.MMShuffle2003 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 1, 0), SimdUtils.Shuffle.MMShuffle2010 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 1, 1), SimdUtils.Shuffle.MMShuffle2011 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 1, 2), SimdUtils.Shuffle.MMShuffle2012 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 1, 3), SimdUtils.Shuffle.MMShuffle2013 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 2, 0), SimdUtils.Shuffle.MMShuffle2020 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 2, 1), SimdUtils.Shuffle.MMShuffle2021 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 2, 2), SimdUtils.Shuffle.MMShuffle2022 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 2, 3), SimdUtils.Shuffle.MMShuffle2023 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 3, 0), SimdUtils.Shuffle.MMShuffle2030 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 3, 1), SimdUtils.Shuffle.MMShuffle2031 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 3, 2), SimdUtils.Shuffle.MMShuffle2032 },
+        { SimdUtils.Shuffle.MMShuffle(2, 0, 3, 3), SimdUtils.Shuffle.MMShuffle2033 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 0, 0), SimdUtils.Shuffle.MMShuffle2100 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 0, 1), SimdUtils.Shuffle.MMShuffle2101 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 0, 2), SimdUtils.Shuffle.MMShuffle2102 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 0, 3), SimdUtils.Shuffle.MMShuffle2103 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 1, 0), SimdUtils.Shuffle.MMShuffle2110 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 1, 1), SimdUtils.Shuffle.MMShuffle2111 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 1, 2), SimdUtils.Shuffle.MMShuffle2112 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 1, 3), SimdUtils.Shuffle.MMShuffle2113 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 2, 0), SimdUtils.Shuffle.MMShuffle2120 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 2, 1), SimdUtils.Shuffle.MMShuffle2121 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 2, 2), SimdUtils.Shuffle.MMShuffle2122 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 2, 3), SimdUtils.Shuffle.MMShuffle2123 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 3, 0), SimdUtils.Shuffle.MMShuffle2130 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 3, 1), SimdUtils.Shuffle.MMShuffle2131 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 3, 2), SimdUtils.Shuffle.MMShuffle2132 },
+        { SimdUtils.Shuffle.MMShuffle(2, 1, 3, 3), SimdUtils.Shuffle.MMShuffle2133 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 0, 0), SimdUtils.Shuffle.MMShuffle2200 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 0, 1), SimdUtils.Shuffle.MMShuffle2201 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 0, 2), SimdUtils.Shuffle.MMShuffle2202 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 0, 3), SimdUtils.Shuffle.MMShuffle2203 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 1, 0), SimdUtils.Shuffle.MMShuffle2210 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 1, 1), SimdUtils.Shuffle.MMShuffle2211 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 1, 2), SimdUtils.Shuffle.MMShuffle2212 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 1, 3), SimdUtils.Shuffle.MMShuffle2213 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 2, 0), SimdUtils.Shuffle.MMShuffle2220 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 2, 1), SimdUtils.Shuffle.MMShuffle2221 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 2, 2), SimdUtils.Shuffle.MMShuffle2222 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 2, 3), SimdUtils.Shuffle.MMShuffle2223 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 3, 0), SimdUtils.Shuffle.MMShuffle2230 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 3, 1), SimdUtils.Shuffle.MMShuffle2231 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 3, 2), SimdUtils.Shuffle.MMShuffle2232 },
+        { SimdUtils.Shuffle.MMShuffle(2, 2, 3, 3), SimdUtils.Shuffle.MMShuffle2233 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 0, 0), SimdUtils.Shuffle.MMShuffle2300 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 0, 1), SimdUtils.Shuffle.MMShuffle2301 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 0, 2), SimdUtils.Shuffle.MMShuffle2302 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 0, 3), SimdUtils.Shuffle.MMShuffle2303 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 1, 0), SimdUtils.Shuffle.MMShuffle2310 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 1, 1), SimdUtils.Shuffle.MMShuffle2311 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 1, 2), SimdUtils.Shuffle.MMShuffle2312 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 1, 3), SimdUtils.Shuffle.MMShuffle2313 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 2, 0), SimdUtils.Shuffle.MMShuffle2320 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 2, 1), SimdUtils.Shuffle.MMShuffle2321 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 2, 2), SimdUtils.Shuffle.MMShuffle2322 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 2, 3), SimdUtils.Shuffle.MMShuffle2323 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 3, 0), SimdUtils.Shuffle.MMShuffle2330 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 3, 1), SimdUtils.Shuffle.MMShuffle2331 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 3, 2), SimdUtils.Shuffle.MMShuffle2332 },
+        { SimdUtils.Shuffle.MMShuffle(2, 3, 3, 3), SimdUtils.Shuffle.MMShuffle2333 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 0, 0), SimdUtils.Shuffle.MMShuffle3000 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 0, 1), SimdUtils.Shuffle.MMShuffle3001 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 0, 2), SimdUtils.Shuffle.MMShuffle3002 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 0, 3), SimdUtils.Shuffle.MMShuffle3003 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 1, 0), SimdUtils.Shuffle.MMShuffle3010 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 1, 1), SimdUtils.Shuffle.MMShuffle3011 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 1, 2), SimdUtils.Shuffle.MMShuffle3012 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 1, 3), SimdUtils.Shuffle.MMShuffle3013 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 2, 0), SimdUtils.Shuffle.MMShuffle3020 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 2, 1), SimdUtils.Shuffle.MMShuffle3021 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 2, 2), SimdUtils.Shuffle.MMShuffle3022 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 2, 3), SimdUtils.Shuffle.MMShuffle3023 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 3, 0), SimdUtils.Shuffle.MMShuffle3030 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 3, 1), SimdUtils.Shuffle.MMShuffle3031 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 3, 2), SimdUtils.Shuffle.MMShuffle3032 },
+        { SimdUtils.Shuffle.MMShuffle(3, 0, 3, 3), SimdUtils.Shuffle.MMShuffle3033 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 0, 0), SimdUtils.Shuffle.MMShuffle3100 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 0, 1), SimdUtils.Shuffle.MMShuffle3101 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 0, 2), SimdUtils.Shuffle.MMShuffle3102 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 0, 3), SimdUtils.Shuffle.MMShuffle3103 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 1, 0), SimdUtils.Shuffle.MMShuffle3110 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 1, 1), SimdUtils.Shuffle.MMShuffle3111 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 1, 2), SimdUtils.Shuffle.MMShuffle3112 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 1, 3), SimdUtils.Shuffle.MMShuffle3113 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 2, 0), SimdUtils.Shuffle.MMShuffle3120 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 2, 1), SimdUtils.Shuffle.MMShuffle3121 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 2, 2), SimdUtils.Shuffle.MMShuffle3122 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 2, 3), SimdUtils.Shuffle.MMShuffle3123 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 3, 0), SimdUtils.Shuffle.MMShuffle3130 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 3, 1), SimdUtils.Shuffle.MMShuffle3131 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 3, 2), SimdUtils.Shuffle.MMShuffle3132 },
+        { SimdUtils.Shuffle.MMShuffle(3, 1, 3, 3), SimdUtils.Shuffle.MMShuffle3133 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 0, 0), SimdUtils.Shuffle.MMShuffle3200 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 0, 1), SimdUtils.Shuffle.MMShuffle3201 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 0, 2), SimdUtils.Shuffle.MMShuffle3202 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 0, 3), SimdUtils.Shuffle.MMShuffle3203 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 1, 0), SimdUtils.Shuffle.MMShuffle3210 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 1, 1), SimdUtils.Shuffle.MMShuffle3211 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 1, 2), SimdUtils.Shuffle.MMShuffle3212 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 1, 3), SimdUtils.Shuffle.MMShuffle3213 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 2, 0), SimdUtils.Shuffle.MMShuffle3220 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 2, 1), SimdUtils.Shuffle.MMShuffle3221 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 2, 2), SimdUtils.Shuffle.MMShuffle3222 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 2, 3), SimdUtils.Shuffle.MMShuffle3223 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 3, 0), SimdUtils.Shuffle.MMShuffle3230 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 3, 1), SimdUtils.Shuffle.MMShuffle3231 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 3, 2), SimdUtils.Shuffle.MMShuffle3232 },
+        { SimdUtils.Shuffle.MMShuffle(3, 2, 3, 3), SimdUtils.Shuffle.MMShuffle3233 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 0, 0), SimdUtils.Shuffle.MMShuffle3300 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 0, 1), SimdUtils.Shuffle.MMShuffle3301 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 0, 2), SimdUtils.Shuffle.MMShuffle3302 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 0, 3), SimdUtils.Shuffle.MMShuffle3303 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 1, 0), SimdUtils.Shuffle.MMShuffle3310 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 1, 1), SimdUtils.Shuffle.MMShuffle3311 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 1, 2), SimdUtils.Shuffle.MMShuffle3312 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 1, 3), SimdUtils.Shuffle.MMShuffle3313 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 2, 0), SimdUtils.Shuffle.MMShuffle3320 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 2, 1), SimdUtils.Shuffle.MMShuffle3321 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 2, 2), SimdUtils.Shuffle.MMShuffle3322 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 2, 3), SimdUtils.Shuffle.MMShuffle3323 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 3, 0), SimdUtils.Shuffle.MMShuffle3330 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 3, 1), SimdUtils.Shuffle.MMShuffle3331 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 3, 2), SimdUtils.Shuffle.MMShuffle3332 },
+        { SimdUtils.Shuffle.MMShuffle(3, 3, 3, 3), SimdUtils.Shuffle.MMShuffle3333 }
+    };
+
+    [Theory]
+    [MemberData(nameof(MMShuffleData))]
+    public void MMShuffleConstantsAreCorrect(byte expected, byte actual)
+        => Assert.Equal(expected, actual);
+
     [Theory]
     [MemberData(nameof(ArraySizesDivisibleBy4))]
     public void BulkShuffleFloat4Channel(int count)
@@ -226,7 +491,7 @@ public partial class SimdUtilsTests
 
         float[] expected = new float[count];
 
-        SimdUtils.Shuffle.InverseMmShuffle(
+        SimdUtils.Shuffle.InverseMMShuffle(
             control,
             out int p3,
             out int p2,
@@ -257,7 +522,7 @@ public partial class SimdUtilsTests
 
         byte[] expected = new byte[count];
 
-        SimdUtils.Shuffle.InverseMmShuffle(
+        SimdUtils.Shuffle.InverseMMShuffle(
             control,
             out int p3,
             out int p2,
@@ -288,7 +553,7 @@ public partial class SimdUtilsTests
 
         byte[] expected = new byte[count];
 
-        SimdUtils.Shuffle.InverseMmShuffle(
+        SimdUtils.Shuffle.InverseMMShuffle(
             control,
             out int _,
             out int p2,
@@ -319,7 +584,7 @@ public partial class SimdUtilsTests
 
         byte[] expected = new byte[result.Length];
 
-        SimdUtils.Shuffle.InverseMmShuffle(
+        SimdUtils.Shuffle.InverseMMShuffle(
             control,
             out int p3,
             out int p2,
@@ -370,7 +635,7 @@ public partial class SimdUtilsTests
 
         byte[] expected = new byte[result.Length];
 
-        SimdUtils.Shuffle.InverseMmShuffle(
+        SimdUtils.Shuffle.InverseMMShuffle(
             control,
             out int _,
             out int p2,