Add initial vectorized color converter implementation

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromCmyk.cs → src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromCmykBasic.cs

@@ -8,16 +8,20 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
 {
     internal abstract partial class JpegColorConverter
     {
-        internal sealed class FromCmyk : JpegColorConverter
+        internal sealed class FromCmykBasic : JpegColorConverter
         {
-            public FromCmyk(int precision)
+            public FromCmykBasic(int precision)
                 : base(JpegColorSpace.Cmyk, precision)
             {
             }
 
             public override void ConvertToRgba(in ComponentValues values, Span<Vector4> result)
             {
-                // TODO: We can optimize a lot here with Vector<float> and SRCS.Unsafe()!
+                ConvertCore(values, result, this.MaximumValue);
+            }
+
+            internal static void ConvertCore(in ComponentValues values, Span<Vector4> result, float maxValue)
+            {
                 ReadOnlySpan<float> cVals = values.Component0;
                 ReadOnlySpan<float> mVals = values.Component1;
                 ReadOnlySpan<float> yVals = values.Component2;
@@ -25,7 +29,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
 
                 var v = new Vector4(0, 0, 0, 1F);
 
-                var maximum = 1 / this.MaximumValue;
+                var maximum = 1 / maxValue;
                 var scale = new Vector4(maximum, maximum, maximum, 1F);
 
                 for (int i = 0; i < result.Length; i++)
@@ -33,7 +37,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
                     float c = cVals[i];
                     float m = mVals[i];
                     float y = yVals[i];
-                    float k = kVals[i] / this.MaximumValue;
+                    float k = kVals[i] / maxValue;
 
                     v.X = c * k;
                     v.Y = m * k;
@@ -47,4 +51,4 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
             }
         }
     }
-}
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromCmykSimdAvx2.cs

@@ -0,0 +1,145 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
+using System;
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+#if SUPPORTS_RUNTIME_INTRINSICS
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+using static SixLabors.ImageSharp.SimdUtils;
+#else
+using SixLabors.ImageSharp.Tuples;
+#endif
+
+namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
+{
+    internal abstract partial class JpegColorConverter
+    {
+        internal sealed class FromCmykVector8 : JpegColorConverter
+        {
+            public FromCmykVector8(int precision)
+                : base(JpegColorSpace.Cmyk, precision)
+            {
+            }
+
+            public static bool IsAvailable => Vector.IsHardwareAccelerated && SimdUtils.HasVector8;
+
+            public override void ConvertToRgba(in ComponentValues values, Span<Vector4> result)
+            {
+                int remainder = result.Length % 8;
+                int simdCount = result.Length - remainder;
+                if (simdCount > 0)
+                {
+                    ConvertCore(values.Slice(0, simdCount), result.Slice(0, simdCount), this.MaximumValue);
+                }
+
+                FromCmykBasic.ConvertCore(values.Slice(simdCount, remainder), result.Slice(simdCount, remainder), this.MaximumValue);
+            }
+
+            internal static void ConvertCore(in ComponentValues values, Span<Vector4> result, float maxValue)
+            {
+                // This implementation is actually AVX specific.
+                // An AVX register is capable of storing 8 float-s.
+                if (!IsAvailable)
+                {
+                    throw new InvalidOperationException(
+                        "JpegColorConverter.FromGrayscaleVector8 can be used only on architecture having 256 byte floating point SIMD registers!");
+                }
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+                ref Vector256<float> cBase =
+                                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
+                ref Vector256<float> mBase =
+                                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
+                ref Vector256<float> yBase =
+                                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
+                ref Vector256<float> kBase =
+                                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3));
+
+                ref Vector256<float> resultBase =
+                    ref Unsafe.As<Vector4, Vector256<float>>(ref MemoryMarshal.GetReference(result));
+
+                // Used for the color conversion
+                var scale = Vector256.Create(1 / maxValue);
+                var one = Vector256.Create(1F);
+
+                // Used for packing
+                ref byte control = ref MemoryMarshal.GetReference(HwIntrinsics.PermuteMaskEvenOdd8x32);
+                Vector256<int> vcontrol = Unsafe.As<byte, Vector256<int>>(ref control);
+
+                int n = result.Length / 8;
+                for (int i = 0; i < n; i++)
+                {
+                    Vector256<float> k = Avx2.PermuteVar8x32(Unsafe.Add(ref kBase, i), vcontrol);
+                    Vector256<float> c = Avx2.PermuteVar8x32(Unsafe.Add(ref cBase, i), vcontrol);
+                    Vector256<float> m = Avx2.PermuteVar8x32(Unsafe.Add(ref mBase, i), vcontrol);
+                    Vector256<float> y = Avx2.PermuteVar8x32(Unsafe.Add(ref yBase, i), vcontrol);
+
+                    k = Avx.Multiply(k, scale);
+
+                    c = Avx.Multiply(Avx.Multiply(c, k), scale);
+                    m = Avx.Multiply(Avx.Multiply(m, k), scale);
+                    y = Avx.Multiply(Avx.Multiply(y, k), scale);
+
+                    Vector256<float> cmLo = Avx.UnpackLow(c, m);
+                    Vector256<float> yoLo = Avx.UnpackLow(y, one);
+                    Vector256<float> cmHi = Avx.UnpackHigh(c, m);
+                    Vector256<float> yoHi = Avx.UnpackHigh(y, one);
+
+                    ref Vector256<float> destination = ref Unsafe.Add(ref resultBase, i * 4);
+
+                    destination = Avx.Shuffle(cmLo, yoLo, 0b01_00_01_00);
+                    Unsafe.Add(ref destination, 1) = Avx.Shuffle(cmLo, yoLo, 0b11_10_11_10);
+                    Unsafe.Add(ref destination, 2) = Avx.Shuffle(cmHi, yoHi, 0b01_00_01_00);
+                    Unsafe.Add(ref destination, 3) = Avx.Shuffle(cmHi, yoHi, 0b11_10_11_10);
+                }
+#else
+                ref Vector<float> cBase =
+                                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component0));
+                ref Vector<float> mBase =
+                                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component1));
+                ref Vector<float> yBase =
+                                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component2));
+                ref Vector<float> kBase =
+                                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component3));
+
+                ref Vector4Octet resultBase =
+                    ref Unsafe.As<Vector4, Vector4Octet>(ref MemoryMarshal.GetReference(result));
+
+                Vector4Pair cc = default;
+                Vector4Pair mm = default;
+                Vector4Pair yy = default;
+                ref Vector<float> ccRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref cc);
+                ref Vector<float> mmRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref mm);
+                ref Vector<float> yyRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref yy);
+
+                var scale = new Vector<float>(1 / maxValue);
+
+                // Walking 8 elements at one step:
+                int n = result.Length / 8;
+                for (int i = 0; i < n; i++)
+                {
+                    Vector<float> c = Unsafe.Add(ref cBase, i);
+                    Vector<float> m = Unsafe.Add(ref mBase, i);
+                    Vector<float> y = Unsafe.Add(ref yBase, i);
+                    Vector<float> k = Unsafe.Add(ref kBase, i) * scale;
+
+                    c = (c * k) * scale;
+                    m = (m * k) * scale;
+                    y = (y * k) * scale;
+
+                    ccRefAsVector = c;
+                    mmRefAsVector = m;
+                    yyRefAsVector = y;
+
+                    // Collect (c0,c1...c8) (m0,m1...m8) (y0,y1...y8) vector values in the expected (r0,g0,g1,1), (r1,g1,g2,1) ... order:
+                    ref Vector4Octet destination = ref Unsafe.Add(ref resultBase, i);
+                    destination.Pack(ref cc, ref mm, ref yy);
+                }
+#endif
+            }
+        }
+    }
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromGrayScale.cs → src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromGrayScaleBasic.cs

@@ -10,16 +10,21 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
 {
     internal abstract partial class JpegColorConverter
     {
-        internal sealed class FromGrayscale : JpegColorConverter
+        internal sealed class FromGrayscaleBasic : JpegColorConverter
         {
-            public FromGrayscale(int precision)
+            public FromGrayscaleBasic(int precision)
                 : base(JpegColorSpace.Grayscale, precision)
             {
             }
 
             public override void ConvertToRgba(in ComponentValues values, Span<Vector4> result)
             {
-                var maximum = 1 / this.MaximumValue;
+                ConvertCore(values, result, this.MaximumValue);
+            }
+
+            internal static void ConvertCore(in ComponentValues values, Span<Vector4> result, float maxValue)
+            {
+                var maximum = 1 / maxValue;
                 var scale = new Vector4(maximum, maximum, maximum, 1F);
 
                 ref float sBase = ref MemoryMarshal.GetReference(values.Component0);
@@ -35,4 +40,4 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
             }
         }
     }
-}
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromGrayScaleSimdAvx2.cs

@@ -0,0 +1,109 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
+using System;
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+#if SUPPORTS_RUNTIME_INTRINSICS
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+using static SixLabors.ImageSharp.SimdUtils;
+#else
+using SixLabors.ImageSharp.Tuples;
+#endif
+
+namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
+{
+    internal abstract partial class JpegColorConverter
+    {
+        internal sealed class FromGrayscaleVector8 : JpegColorConverter
+        {
+            public FromGrayscaleVector8(int precision)
+                : base(JpegColorSpace.Grayscale, precision)
+            {
+            }
+
+            public static bool IsAvailable => Vector.IsHardwareAccelerated && SimdUtils.HasVector8;
+
+            public override void ConvertToRgba(in ComponentValues values, Span<Vector4> result)
+            {
+                int remainder = result.Length % 8;
+                int simdCount = result.Length - remainder;
+                if (simdCount > 0)
+                {
+                    ConvertCore(values.Slice(0, simdCount), result.Slice(0, simdCount), this.MaximumValue);
+                }
+
+                FromGrayscaleBasic.ConvertCore(values.Slice(simdCount, remainder), result.Slice(simdCount, remainder), this.MaximumValue);
+            }
+
+            internal static void ConvertCore(in ComponentValues values, Span<Vector4> result, float maxValue)
+            {
+                // This implementation is actually AVX specific.
+                // An AVX register is capable of storing 8 float-s.
+                if (!IsAvailable)
+                {
+                    throw new InvalidOperationException(
+                        "JpegColorConverter.FromGrayscaleVector8 can be used only on architecture having 256 byte floating point SIMD registers!");
+                }
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+                ref Vector256<float> gBase =
+                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
+
+                ref Vector256<float> resultBase =
+                    ref Unsafe.As<Vector4, Vector256<float>>(ref MemoryMarshal.GetReference(result));
+
+                // Used for the color conversion
+                var scale = Vector256.Create(1 / maxValue);
+                var one = Vector256.Create(1F);
+
+                // Used for packing
+                ref byte control = ref MemoryMarshal.GetReference(HwIntrinsics.PermuteMaskEvenOdd8x32);
+                Vector256<int> vcontrol = Unsafe.As<byte, Vector256<int>>(ref control);
+
+                int n = result.Length / 8;
+                for (int i = 0; i < n; i++)
+                {
+                    Vector256<float> g = Avx2.PermuteVar8x32(Unsafe.Add(ref gBase, i), vcontrol);
+
+                    g = Avx.Multiply(g, scale);
+
+                    ref Vector256<float> destination = ref Unsafe.Add(ref resultBase, i * 4);
+
+                    destination = Avx.Blend(Avx.Permute(g, 0b00_00_00_00), one, 0b1000_1000);
+                    Unsafe.Add(ref destination, 1) = Avx.Blend(Avx.Permute(g, 0b01_01_01_01), one, 0b1000_1000);
+                    Unsafe.Add(ref destination, 2) = Avx.Blend(Avx.Permute(g, 0b10_10_10_10), one, 0b1000_1000);
+                    Unsafe.Add(ref destination, 3) = Avx.Blend(Avx.Permute(g, 0b11_11_11_11), one, 0b1000_1000);
+                }
+#else
+                ref Vector<float> gBase =
+                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component0));
+
+                ref Vector4Octet resultBase =
+                    ref Unsafe.As<Vector4, Vector4Octet>(ref MemoryMarshal.GetReference(result));
+
+                Vector4Pair gg = default;
+                ref Vector<float> ggRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref gg);
+
+                var scale = new Vector<float>(1 / maxValue);
+
+                // Walking 8 elements at one step:
+                int n = result.Length / 8;
+                for (int i = 0; i < n; i++)
+                {
+                    Vector<float> g = Unsafe.Add(ref gBase, i);
+                    g *= scale;
+
+                    ggRefAsVector = g;
+
+                    // Collect (g0,g1...g7) vector values in the expected (g0,g0,g0,1), (g1,g1,g1,1) ... order:
+                    ref Vector4Octet destination = ref Unsafe.Add(ref resultBase, i);
+                    destination.Pack(ref gg);
+                }
+#endif
+            }
+        }
+    }
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromRgb.cs → src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromRgbBasic.cs

@@ -8,23 +8,27 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
 {
     internal abstract partial class JpegColorConverter
     {
-        internal sealed class FromRgb : JpegColorConverter
+        internal sealed class FromRgbBasic : JpegColorConverter
         {
-            public FromRgb(int precision)
+            public FromRgbBasic(int precision)
                 : base(JpegColorSpace.RGB, precision)
             {
             }
 
             public override void ConvertToRgba(in ComponentValues values, Span<Vector4> result)
             {
-                // TODO: We can optimize a lot here with Vector<float> and SRCS.Unsafe()!
+                ConvertCore(values, result, this.MaximumValue);
+            }
+
+            internal static void ConvertCore(in ComponentValues values, Span<Vector4> result, float maxValue)
+            {
                 ReadOnlySpan<float> rVals = values.Component0;
                 ReadOnlySpan<float> gVals = values.Component1;
                 ReadOnlySpan<float> bVals = values.Component2;
 
                 var v = new Vector4(0, 0, 0, 1);
 
-                var maximum = 1 / this.MaximumValue;
+                var maximum = 1 / maxValue;
                 var scale = new Vector4(maximum, maximum, maximum, 1F);
 
                 for (int i = 0; i < result.Length; i++)
@@ -44,4 +48,4 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
             }
         }
     }
-}
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromRgbSimdAvx2.cs

@@ -0,0 +1,132 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
+using System;
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+#if SUPPORTS_RUNTIME_INTRINSICS
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+using static SixLabors.ImageSharp.SimdUtils;
+#else
+using SixLabors.ImageSharp.Tuples;
+#endif
+
+namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
+{
+    internal abstract partial class JpegColorConverter
+    {
+        internal sealed class FromRgbVector8 : JpegColorConverter
+        {
+            public FromRgbVector8(int precision)
+                : base(JpegColorSpace.RGB, precision)
+            {
+            }
+
+            public static bool IsAvailable => Vector.IsHardwareAccelerated && SimdUtils.HasVector8;
+
+            public override void ConvertToRgba(in ComponentValues values, Span<Vector4> result)
+            {
+                int remainder = result.Length % 8;
+                int simdCount = result.Length - remainder;
+                if (simdCount > 0)
+                {
+                    ConvertCore(values.Slice(0, simdCount), result.Slice(0, simdCount), this.MaximumValue);
+                }
+
+                FromRgbBasic.ConvertCore(values.Slice(simdCount, remainder), result.Slice(simdCount, remainder), this.MaximumValue);
+            }
+
+            internal static void ConvertCore(in ComponentValues values, Span<Vector4> result, float maxValue)
+            {
+                // This implementation is actually AVX specific.
+                // An AVX register is capable of storing 8 float-s.
+                if (!IsAvailable)
+                {
+                    throw new InvalidOperationException(
+                        "JpegColorConverter.FromGrayscaleVector8 can be used only on architecture having 256 byte floating point SIMD registers!");
+                }
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+                ref Vector256<float> rBase =
+                                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
+                ref Vector256<float> gBase =
+                                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
+                ref Vector256<float> bBase =
+                                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
+
+                ref Vector256<float> resultBase =
+                    ref Unsafe.As<Vector4, Vector256<float>>(ref MemoryMarshal.GetReference(result));
+
+                // Used for the color conversion
+                var scale = Vector256.Create(1 / maxValue);
+                var one = Vector256.Create(1F);
+
+                // Used for packing
+                ref byte control = ref MemoryMarshal.GetReference(HwIntrinsics.PermuteMaskEvenOdd8x32);
+                Vector256<int> vcontrol = Unsafe.As<byte, Vector256<int>>(ref control);
+
+                int n = result.Length / 8;
+                for (int i = 0; i < n; i++)
+                {
+                    Vector256<float> r = Avx.Multiply(Avx2.PermuteVar8x32(Unsafe.Add(ref rBase, i), vcontrol), scale);
+                    Vector256<float> g = Avx.Multiply(Avx2.PermuteVar8x32(Unsafe.Add(ref gBase, i), vcontrol), scale);
+                    Vector256<float> b = Avx.Multiply(Avx2.PermuteVar8x32(Unsafe.Add(ref bBase, i), vcontrol), scale);
+
+                    Vector256<float> rgLo = Avx.UnpackLow(r, g);
+                    Vector256<float> boLo = Avx.UnpackLow(b, one);
+                    Vector256<float> rgHi = Avx.UnpackHigh(r, g);
+                    Vector256<float> boHi = Avx.UnpackHigh(b, one);
+
+                    ref Vector256<float> destination = ref Unsafe.Add(ref resultBase, i * 4);
+
+                    destination = Avx.Shuffle(rgLo, boLo, 0b01_00_01_00);
+                    Unsafe.Add(ref destination, 1) = Avx.Shuffle(rgLo, boLo, 0b11_10_11_10);
+                    Unsafe.Add(ref destination, 2) = Avx.Shuffle(rgHi, boHi, 0b01_00_01_00);
+                    Unsafe.Add(ref destination, 3) = Avx.Shuffle(rgHi, boHi, 0b11_10_11_10);
+                }
+#else
+                ref Vector<float> rBase =
+                                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component0));
+                ref Vector<float> gBase =
+                                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component1));
+                ref Vector<float> bBase =
+                                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component2));
+
+                ref Vector4Octet resultBase =
+                    ref Unsafe.As<Vector4, Vector4Octet>(ref MemoryMarshal.GetReference(result));
+
+                Vector4Pair rr = default;
+                Vector4Pair gg = default;
+                Vector4Pair bb = default;
+                ref Vector<float> rrRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref rr);
+                ref Vector<float> ggRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref gg);
+                ref Vector<float> bbRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref bb);
+
+                var scale = new Vector<float>(1 / maxValue);
+
+                // Walking 8 elements at one step:
+                int n = result.Length / 8;
+                for (int i = 0; i < n; i++)
+                {
+                    Vector<float> r = Unsafe.Add(ref rBase, i);
+                    Vector<float> g = Unsafe.Add(ref gBase, i);
+                    Vector<float> b = Unsafe.Add(ref bBase, i);
+                    r *= scale;
+                    g *= scale;
+                    b *= scale;
+
+                    rrRefAsVector = r;
+                    ggRefAsVector = g;
+                    bbRefAsVector = b;
+
+                    // Collect (r0,r1...r8) (g0,g1...g8) (b0,b1...b8) vector values in the expected (r0,g0,g1,1), (r1,g1,g2,1) ... order:
+                    ref Vector4Octet destination = ref Unsafe.Add(ref resultBase, i);
+                    destination.Pack(ref rr, ref gg, ref bb);
+                }
+#endif
+            }
+        }
+    }
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromYCbCrSimdAvx2.cs

@@ -9,8 +9,9 @@ using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
 using System.Runtime.Intrinsics.X86;
 using static SixLabors.ImageSharp.SimdUtils;
-#endif
+#else
 using SixLabors.ImageSharp.Tuples;
+#endif
 
 // ReSharper disable ImpureMethodCallOnReadonlyValueField
 namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
@@ -98,7 +99,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
                     Vector256<float> g = HwIntrinsics.MultiplyAdd(HwIntrinsics.MultiplyAdd(y, cb, gCbMult), cr, gCrMult);
                     Vector256<float> b = HwIntrinsics.MultiplyAdd(y, cb, bCbMult);
 
-                    // TODO: We should be savving to RGBA not Vector4
+                    // TODO: We should be saving to RGBA not Vector4
                     r = Avx.Multiply(Avx.RoundToNearestInteger(r), scale);
                     g = Avx.Multiply(Avx.RoundToNearestInteger(g), scale);
                     b = Avx.Multiply(Avx.RoundToNearestInteger(b), scale);

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromYccK.cs → src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromYccKBasic.cs

@@ -8,14 +8,19 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
 {
     internal abstract partial class JpegColorConverter
     {
-        internal sealed class FromYccK : JpegColorConverter
+        internal sealed class FromYccKBasic : JpegColorConverter
         {
-            public FromYccK(int precision)
+            public FromYccKBasic(int precision)
                 : base(JpegColorSpace.Ycck, precision)
             {
             }
 
             public override void ConvertToRgba(in ComponentValues values, Span<Vector4> result)
+            {
+                ConvertCore(values, result, this.MaximumValue, this.HalfValue);
+            }
+
+            internal static void ConvertCore(in ComponentValues values, Span<Vector4> result, float maxValue, float halfValue)
             {
                 // TODO: We can optimize a lot here with Vector<float> and SRCS.Unsafe()!
                 ReadOnlySpan<float> yVals = values.Component0;
@@ -25,19 +30,19 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
 
                 var v = new Vector4(0, 0, 0, 1F);
 
-                var maximum = 1 / this.MaximumValue;
+                var maximum = 1 / maxValue;
                 var scale = new Vector4(maximum, maximum, maximum, 1F);
 
                 for (int i = 0; i < result.Length; i++)
                 {
                     float y = yVals[i];
-                    float cb = cbVals[i] - this.HalfValue;
-                    float cr = crVals[i] - this.HalfValue;
-                    float k = kVals[i] / this.MaximumValue;
+                    float cb = cbVals[i] - halfValue;
+                    float cr = crVals[i] - halfValue;
+                    float k = kVals[i] / maxValue;
 
-                    v.X = (this.MaximumValue - MathF.Round(y + (1.402F * cr), MidpointRounding.AwayFromZero)) * k;
-                    v.Y = (this.MaximumValue - MathF.Round(y - (0.344136F * cb) - (0.714136F * cr), MidpointRounding.AwayFromZero)) * k;
-                    v.Z = (this.MaximumValue - MathF.Round(y + (1.772F * cb), MidpointRounding.AwayFromZero)) * k;
+                    v.X = (maxValue - MathF.Round(y + (1.402F * cr), MidpointRounding.AwayFromZero)) * k;
+                    v.Y = (maxValue - MathF.Round(y - (0.344136F * cb) - (0.714136F * cr), MidpointRounding.AwayFromZero)) * k;
+                    v.Z = (maxValue - MathF.Round(y + (1.772F * cb), MidpointRounding.AwayFromZero)) * k;
                     v.W = 1F;
 
                     v *= scale;
@@ -47,4 +52,4 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
             }
         }
     }
-}
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromYccKSimdAvx2.cs

@@ -0,0 +1,193 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
+using System;
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+#if SUPPORTS_RUNTIME_INTRINSICS
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+using static SixLabors.ImageSharp.SimdUtils;
+#else
+using SixLabors.ImageSharp.Tuples;
+#endif
+
+namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
+{
+    internal abstract partial class JpegColorConverter
+    {
+        internal sealed class FromYccKVector8 : JpegColorConverter
+        {
+            public FromYccKVector8(int precision)
+                : base(JpegColorSpace.Ycck, precision)
+            {
+            }
+
+            public static bool IsAvailable => Vector.IsHardwareAccelerated && SimdUtils.HasVector8;
+
+            public override void ConvertToRgba(in ComponentValues values, Span<Vector4> result)
+            {
+                int remainder = result.Length % 8;
+                int simdCount = result.Length - remainder;
+                if (simdCount > 0)
+                {
+                    ConvertCore(values.Slice(0, simdCount), result.Slice(0, simdCount), this.MaximumValue, this.HalfValue);
+                }
+
+                FromYccKBasic.ConvertCore(values.Slice(simdCount, remainder), result.Slice(simdCount, remainder), this.MaximumValue, this.HalfValue);
+            }
+
+            internal static void ConvertCore(in ComponentValues values, Span<Vector4> result, float maxValue, float halfValue)
+            {
+                // This implementation is actually AVX specific.
+                // An AVX register is capable of storing 8 float-s.
+                if (!IsAvailable)
+                {
+                    throw new InvalidOperationException(
+                        "JpegColorConverter.FromYCbCrSimd256 can be used only on architecture having 256 byte floating point SIMD registers!");
+                }
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+                ref Vector256<float> yBase =
+                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0));
+                ref Vector256<float> cbBase =
+                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1));
+                ref Vector256<float> crBase =
+                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2));
+                ref Vector256<float> kBase =
+                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3));
+
+                ref Vector256<float> resultBase =
+                    ref Unsafe.As<Vector4, Vector256<float>>(ref MemoryMarshal.GetReference(result));
+
+                // Used for the color conversion
+                var chromaOffset = Vector256.Create(-halfValue);
+                var scale = Vector256.Create(1 / maxValue);
+                var max = Vector256.Create(maxValue);
+                var rCrMult = Vector256.Create(1.402F);
+                var gCbMult = Vector256.Create(-0.344136F);
+                var gCrMult = Vector256.Create(-0.714136F);
+                var bCbMult = Vector256.Create(1.772F);
+
+                // Used for packing.
+                var va = Vector256.Create(1F);
+                ref byte control = ref MemoryMarshal.GetReference(HwIntrinsics.PermuteMaskEvenOdd8x32);
+                Vector256<int> vcontrol = Unsafe.As<byte, Vector256<int>>(ref control);
+
+                // Walking 8 elements at one step:
+                int n = result.Length / 8;
+                for (int i = 0; i < n; i++)
+                {
+                    // y = yVals[i];
+                    // cb = cbVals[i] - 128F;
+                    // cr = crVals[i] - 128F;
+                    // k = kVals[i] / 256F;
+                    Vector256<float> y = Unsafe.Add(ref yBase, i);
+                    Vector256<float> cb = Avx.Add(Unsafe.Add(ref cbBase, i), chromaOffset);
+                    Vector256<float> cr = Avx.Add(Unsafe.Add(ref crBase, i), chromaOffset);
+                    Vector256<float> k = Avx.Divide(Unsafe.Add(ref kBase, i), max);
+
+                    y = Avx2.PermuteVar8x32(y, vcontrol);
+                    cb = Avx2.PermuteVar8x32(cb, vcontrol);
+                    cr = Avx2.PermuteVar8x32(cr, vcontrol);
+                    k = Avx2.PermuteVar8x32(k, vcontrol);
+
+                    // r = y + (1.402F * cr);
+                    // g = y - (0.344136F * cb) - (0.714136F * cr);
+                    // b = y + (1.772F * cb);
+                    // Adding & multiplying 8 elements at one time:
+                    Vector256<float> r = HwIntrinsics.MultiplyAdd(y, cr, rCrMult);
+                    Vector256<float> g = HwIntrinsics.MultiplyAdd(HwIntrinsics.MultiplyAdd(y, cb, gCbMult), cr, gCrMult);
+                    Vector256<float> b = HwIntrinsics.MultiplyAdd(y, cb, bCbMult);
+
+                    r = Avx.Subtract(max, Avx.RoundToNearestInteger(r));
+                    g = Avx.Subtract(max, Avx.RoundToNearestInteger(g));
+                    b = Avx.Subtract(max, Avx.RoundToNearestInteger(b));
+
+                    r = Avx.Multiply(Avx.Multiply(r, k), scale);
+                    g = Avx.Multiply(Avx.Multiply(g, k), scale);
+                    b = Avx.Multiply(Avx.Multiply(b, k), scale);
+
+                    Vector256<float> vte = Avx.UnpackLow(r, b);
+                    Vector256<float> vto = Avx.UnpackLow(g, va);
+
+                    ref Vector256<float> destination = ref Unsafe.Add(ref resultBase, i * 4);
+
+                    destination = Avx.UnpackLow(vte, vto);
+                    Unsafe.Add(ref destination, 1) = Avx.UnpackHigh(vte, vto);
+
+                    vte = Avx.UnpackHigh(r, b);
+                    vto = Avx.UnpackHigh(g, va);
+
+                    Unsafe.Add(ref destination, 2) = Avx.UnpackLow(vte, vto);
+                    Unsafe.Add(ref destination, 3) = Avx.UnpackHigh(vte, vto);
+                }
+#else
+                ref Vector<float> yBase =
+                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component0));
+                ref Vector<float> cbBase =
+                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component1));
+                ref Vector<float> crBase =
+                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component2));
+                ref Vector<float> kBase =
+                    ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(values.Component3));
+
+                ref Vector4Octet resultBase =
+                    ref Unsafe.As<Vector4, Vector4Octet>(ref MemoryMarshal.GetReference(result));
+
+                var chromaOffset = new Vector<float>(-halfValue);
+
+                // Walking 8 elements at one step:
+                int n = result.Length / 8;
+
+                Vector4Pair rr = default;
+                Vector4Pair gg = default;
+                Vector4Pair bb = default;
+
+                ref Vector<float> rrRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref rr);
+                ref Vector<float> ggRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref gg);
+                ref Vector<float> bbRefAsVector = ref Unsafe.As<Vector4Pair, Vector<float>>(ref bb);
+
+                var scale = new Vector<float>(1 / maxValue);
+                var max = new Vector<float>(maxValue);
+
+                for (int i = 0; i < n; i++)
+                {
+                    // y = yVals[i];
+                    // cb = cbVals[i] - 128F;
+                    // cr = crVals[i] - 128F;
+                    // k = kVals[i] / 256F;
+                    Vector<float> y = Unsafe.Add(ref yBase, i);
+                    Vector<float> cb = Unsafe.Add(ref cbBase, i) + chromaOffset;
+                    Vector<float> cr = Unsafe.Add(ref crBase, i) + chromaOffset;
+                    Vector<float> k = Unsafe.Add(ref kBase, i) / max;
+
+                    // r = y + (1.402F * cr);
+                    // g = y - (0.344136F * cb) - (0.714136F * cr);
+                    // b = y + (1.772F * cb);
+                    // Adding & multiplying 8 elements at one time:
+                    Vector<float> r = y + (cr * new Vector<float>(1.402F));
+                    Vector<float> g = y - (cb * new Vector<float>(0.344136F)) - (cr * new Vector<float>(0.714136F));
+                    Vector<float> b = y + (cb * new Vector<float>(1.772F));
+
+                    r = (max - r.FastRound()) * k;
+                    g = (max - g.FastRound()) * k;
+                    b = (max - b.FastRound()) * k;
+                    r *= scale;
+                    g *= scale;
+                    b *= scale;
+
+                    rrRefAsVector = r;
+                    ggRefAsVector = g;
+                    bbRefAsVector = b;
+
+                    // Collect (r0,r1...r8) (g0,g1...g8) (b0,b1...b8) vector values in the expected (r0,g0,g1,1), (r1,g1,g2,1) ... order:
+                    ref Vector4Octet destination = ref Unsafe.Add(ref resultBase, i);
+                    destination.Pack(ref rr, ref gg, ref bb);
+                }
+#endif
+            }
+        }
+    }
+}

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

@@ -21,17 +21,17 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
             {
                 // 8-bit converters
                 GetYCbCrConverter(8),
-                new FromYccK(8),
-                new FromCmyk(8),
-                new FromGrayscale(8),
-                new FromRgb(8),
+                GetYccKConverter(8),
+                GetCmykConverter(8),
+                GetGrayScaleConverter(8),
+                GetRgbConverter(8),
 
                 // 12-bit converters
                 GetYCbCrConverter(12),
-                new FromYccK(12),
-                new FromCmyk(12),
-                new FromGrayscale(12),
-                new FromRgb(12),
+                GetYccKConverter(12),
+                GetCmykConverter(12),
+                GetGrayScaleConverter(12),
+                GetRgbConverter(12),
             };
 
         /// <summary>
@@ -94,6 +94,30 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
         private static JpegColorConverter GetYCbCrConverter(int precision) =>
             FromYCbCrSimdVector8.IsAvailable ? (JpegColorConverter)new FromYCbCrSimdVector8(precision) : new FromYCbCrSimd(precision);
 
+        /// <summary>
+        /// Returns the <see cref="JpegColorConverter"/> for the YccK colorspace that matches the current CPU architecture.
+        /// </summary>
+        private static JpegColorConverter GetYccKConverter(int precision) =>
+            FromYccKVector8.IsAvailable ? (JpegColorConverter)new FromYccKVector8(precision) : new FromYccKBasic(precision);
+
+        /// <summary>
+        /// Returns the <see cref="JpegColorConverter"/> for the CMYK colorspace that matches the current CPU architecture.
+        /// </summary>
+        private static JpegColorConverter GetCmykConverter(int precision) =>
+            FromCmykVector8.IsAvailable ? (JpegColorConverter)new FromCmykVector8(precision) : new FromCmykBasic(precision);
+
+        /// <summary>
+        /// Returns the <see cref="JpegColorConverter"/> for the gray scale colorspace that matches the current CPU architecture.
+        /// </summary>
+        private static JpegColorConverter GetGrayScaleConverter(int precision) =>
+            FromGrayscaleVector8.IsAvailable ? (JpegColorConverter)new FromGrayscaleVector8(precision) : new FromGrayscaleBasic(precision);
+
+        /// <summary>
+        /// Returns the <see cref="JpegColorConverter"/> for the RGB colorspace that matches the current CPU architecture.
+        /// </summary>
+        private static JpegColorConverter GetRgbConverter(int precision) =>
+            FromRgbVector8.IsAvailable ? (JpegColorConverter)new FromRgbVector8(precision) : new FromRgbBasic(precision);
+
         /// <summary>
         /// A stack-only struct to reference the input buffers using <see cref="ReadOnlySpan{T}"/>-s.
         /// </summary>
@@ -229,6 +253,52 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
                 this.V7.Z = b.B.W;
                 this.V7.W = 1f;
             }
+
+            /// <summary>
+            /// Pack (g0,g1...g7) vector values as (g0,g0,g0,1), (g1,g1,g1,1) ...
+            /// </summary>
+            public void Pack(ref Vector4Pair g)
+            {
+                this.V0.X = g.A.X;
+                this.V0.Y = g.A.X;
+                this.V0.Z = g.A.X;
+                this.V0.W = 1f;
+
+                this.V1.X = g.A.Y;
+                this.V1.Y = g.A.Y;
+                this.V1.Z = g.A.Y;
+                this.V1.W = 1f;
+
+                this.V2.X = g.A.Z;
+                this.V2.Y = g.A.Z;
+                this.V2.Z = g.A.Z;
+                this.V2.W = 1f;
+
+                this.V3.X = g.A.W;
+                this.V3.Y = g.A.W;
+                this.V3.Z = g.A.W;
+                this.V3.W = 1f;
+
+                this.V4.X = g.B.X;
+                this.V4.Y = g.B.X;
+                this.V4.Z = g.B.X;
+                this.V4.W = 1f;
+
+                this.V5.X = g.B.Y;
+                this.V5.Y = g.B.Y;
+                this.V5.Z = g.B.Y;
+                this.V5.W = 1f;
+
+                this.V6.X = g.B.Z;
+                this.V6.Y = g.B.Z;
+                this.V6.Z = g.B.Z;
+                this.V6.W = 1f;
+
+                this.V7.X = g.B.W;
+                this.V7.Y = g.B.W;
+                this.V7.Z = g.B.W;
+                this.V7.W = 1f;
+            }
         }
     }
 }

tests/ImageSharp.Tests/Formats/Jpg/JpegColorConverterTests.cs

@@ -183,7 +183,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 var actual = new Rgb(rgba.X, rgba.Y, rgba.Z);
                 var expected = new Rgb(v.X, v.Y, v.Z);
 
-                Assert.Equal(expected, actual);
+                Assert.Equal(expected, actual, ColorSpaceComparer);
                 Assert.Equal(1, rgba.W);
             }
         }