JpegColorConverter.FromYCbCrSimd256 conversion core method seems to work

src/ImageSharp/Common/Extensions/Vector4Extensions.cs

@@ -79,5 +79,16 @@ namespace SixLabors.ImageSharp
 
             return MathF.Pow((signal + 0.055F) / 1.055F, 2.4F);
         }
+
+        /// <summary>
+        /// Transform all scalars in 'v' in a way that converting them to <see cref="int"/> would have rounding semantics.
+        /// </summary>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        internal static Vector4 PseudoRound(this Vector4 v)
+        {
+            var sign = Vector4.Clamp(v, new Vector4(-1), new Vector4(1));
+
+            return v + (sign * 0.5f);
+        }
     }
 }

src/ImageSharp/Formats/Jpeg/Common/Decoder/JpegColorConverter.FromCmyk.cs

@@ -5,7 +5,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
 {
     internal abstract partial class JpegColorConverter
     {
-        private class FromCmyk : JpegColorConverter
+        internal class FromCmyk : JpegColorConverter
         {
             public FromCmyk()
                 : base(JpegColorSpace.Cmyk)

src/ImageSharp/Formats/Jpeg/Common/Decoder/JpegColorConverter.FromGrayScale.cs

@@ -5,7 +5,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
 {
     internal abstract partial class JpegColorConverter
     {
-        private class FromGrayScale : JpegColorConverter
+        internal class FromGrayScale : JpegColorConverter
         {
             public FromGrayScale()
                 : base(JpegColorSpace.GrayScale)

src/ImageSharp/Formats/Jpeg/Common/Decoder/JpegColorConverter.FromRgb.cs

@@ -5,7 +5,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
 {
     internal abstract partial class JpegColorConverter
     {
-        private class FromRgb : JpegColorConverter
+        internal class FromRgb : JpegColorConverter
         {
             public FromRgb()
                 : base(JpegColorSpace.RGB)

src/ImageSharp/Formats/Jpeg/Common/Decoder/JpegColorConverter.FromYCbCr.cs

@@ -1,18 +1,26 @@
 using System;
 using System.Numerics;
+using System.Runtime.CompilerServices;
+
+// ReSharper disable InconsistentNaming
 
 namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
 {
     internal abstract partial class JpegColorConverter
     {
-        private class FromYCbCr : JpegColorConverter
+        internal class FromYCbCrBasic : JpegColorConverter
         {
-            public FromYCbCr()
+            public FromYCbCrBasic()
                 : base(JpegColorSpace.YCbCr)
             {
             }
 
             public override void ConvertToRGBA(ComponentValues values, Span<Vector4> result)
+            {
+                ConvertCore(values, result);
+            }
+
+            internal static void ConvertCore(ComponentValues values, Span<Vector4> result)
             {
                 // TODO: We can optimize a lot here with Vector<float> and SRCS.Unsafe()!
                 ReadOnlySpan<float> yVals = values.Component0;
@@ -39,5 +47,154 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
                 }
             }
         }
+
+        internal class FromYCbCrSimd256 : JpegColorConverter
+        {
+            public FromYCbCrSimd256()
+                : base(JpegColorSpace.YCbCr)
+            {
+            }
+
+            public override void ConvertToRGBA(ComponentValues values, Span<Vector4> result)
+            {   
+            }
+
+            private struct Vector4Pair
+            {
+                public Vector4 A;
+
+                public Vector4 B;
+
+                private static readonly Vector4 Scale = new Vector4(1 / 255F);
+
+                [MethodImpl(MethodImplOptions.AggressiveInlining)]
+                public void RoundAndDownscale()
+                {
+                    this.A = this.A.PseudoRound() * Scale;
+                    this.B = this.B.PseudoRound() * Scale;
+                }
+            }
+
+            private struct Vector4Octet
+            {
+#pragma warning disable SA1132 // Do not combine fields
+                public Vector4 V0, V1, V2, V3, V4, V5, V6, V7;
+#pragma warning restore SA1132 // Do not combine fields
+
+                public static Vector4Octet CreateCollector()
+                {
+                    var result = default(Vector4Octet);
+                    result.V0.W = 1f;
+                    result.V1.W = 1f;
+                    result.V2.W = 1f;
+                    result.V3.W = 1f;
+                    result.V4.W = 1f;
+                    result.V5.W = 1f;
+                    result.V6.W = 1f;
+                    result.V7.W = 1f;
+                    return result;
+                }
+
+                public void Collect(ref Vector4Pair rr, ref Vector4Pair gg, ref Vector4Pair bb)
+                {
+                    this.V0.X = rr.A.X;
+                    this.V0.Y = gg.A.X;
+                    this.V0.Z = bb.A.X;
+                    this.V0.W = 1f;
+
+                    this.V1.X = rr.A.Y;
+                    this.V1.Y = gg.A.Y;
+                    this.V1.Z = bb.A.Y;
+                    this.V1.W = 1f;
+
+                    this.V2.X = rr.A.Z;
+                    this.V2.Y = gg.A.Z;
+                    this.V2.Z = bb.A.Z;
+                    this.V2.W = 1f;
+
+                    this.V3.X = rr.A.W;
+                    this.V3.Y = gg.A.W;
+                    this.V3.Z = bb.A.W;
+                    this.V3.W = 1f;
+
+                    this.V4.X = rr.B.X;
+                    this.V4.Y = gg.B.X;
+                    this.V4.Z = bb.B.X;
+                    this.V4.W = 1f;
+
+                    this.V5.X = rr.B.Y;
+                    this.V5.Y = gg.B.Y;
+                    this.V5.Z = bb.B.Y;
+                    this.V5.W = 1f;
+
+                    this.V6.X = rr.B.Z;
+                    this.V6.Y = gg.B.Z;
+                    this.V6.Z = bb.B.Z;
+                    this.V6.W = 1f;
+
+                    this.V7.X = rr.B.W;
+                    this.V7.Y = gg.B.W;
+                    this.V7.Z = bb.B.W;
+                    this.V7.W = 1f;
+                }
+            }
+
+            internal static void ConvertAligned(ComponentValues values, Span<Vector4> result)
+            {
+                if (!IsAvailable)
+                {
+                    throw new InvalidOperationException(
+                        "JpegColorConverter.FromYCbCrSimd256 can be used only on architecture having 256 byte floating point SIMD registers!");
+                }
+
+                ref Vector<float> yBase =
+                    ref Unsafe.As<float, Vector<float>>(ref values.Component0.DangerousGetPinnableReference());
+                ref Vector<float> cbBase =
+                    ref Unsafe.As<float, Vector<float>>(ref values.Component1.DangerousGetPinnableReference());
+                ref Vector<float> crBase =
+                    ref Unsafe.As<float, Vector<float>>(ref values.Component2.DangerousGetPinnableReference());
+
+                ref Vector4Octet resultBase =
+                    ref Unsafe.As<Vector4, Vector4Octet>(ref result.DangerousGetPinnableReference());
+
+                var chromaOffset = new Vector<float>(-128f);
+
+                int n = result.Length / 8;
+
+                for (int i = 0; i < n; i++)
+                {
+                    // y = yVals[i];
+                    // cb = cbVals[i] - 128F;
+                    // cr = crVals[i] - 128F;
+                    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;
+
+                    // 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));
+
+                    // Vector<float> has no .Clamp(), need to switch to Vector4 for the next operation:
+                    // TODO: Is it worth to use Vector<float> at all?
+                    Vector4Pair rr = Unsafe.As<Vector<float>, Vector4Pair>(ref r);
+                    Vector4Pair gg = Unsafe.As<Vector<float>, Vector4Pair>(ref g);
+                    Vector4Pair bb = Unsafe.As<Vector<float>, Vector4Pair>(ref b);
+
+                    rr.RoundAndDownscale();
+                    gg.RoundAndDownscale();
+                    bb.RoundAndDownscale();
+
+                    // 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.Collect(ref rr, ref gg, ref bb);
+                }
+            }
+
+            public static bool IsAvailable => Vector<float>.Count == 8;
+        }
     }
 }

src/ImageSharp/Formats/Jpeg/Common/Decoder/JpegColorConverter.FromYccK.cs

@@ -5,7 +5,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
 {
     internal abstract partial class JpegColorConverter
     {
-        private class FromYccK : JpegColorConverter
+        internal class FromYccK : JpegColorConverter
         {
             public FromYccK()
                 : base(JpegColorSpace.Ycck)

src/ImageSharp/Formats/Jpeg/Common/Decoder/JpegColorConverter.cs

@@ -15,7 +15,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
         /// <summary>
         /// The avalilable converters
         /// </summary>
-        private static readonly JpegColorConverter[] Converters = { new FromYCbCr(), new FromYccK(), new FromCmyk(), new FromGrayScale(), new FromRgb() };
+        private static readonly JpegColorConverter[] Converters = { new FromYCbCrBasic(), new FromYccK(), new FromCmyk(), new FromGrayScale(), new FromRgb() };
 
         /// <summary>
         /// Initializes a new instance of the <see cref="JpegColorConverter"/> class.

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

@@ -1,5 +1,3 @@
-namespace SixLabors.ImageSharp.Tests.Formats.Jpg
-{
 using System;
 using System.Numerics;
 
@@ -10,10 +8,13 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
 using Xunit;
 using Xunit.Abstractions;
+// ReSharper disable InconsistentNaming
 
+namespace SixLabors.ImageSharp.Tests.Formats.Jpg
+{
     public class JpegColorConverterTests
     {
-        private const float Precision = 0.01f;
+        private const float Precision = 1/255f;
 
         public static readonly TheoryData<int, int, int> CommonConversionData =
             new TheoryData<int, int, int>
@@ -32,17 +33,44 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
         private ITestOutputHelper Output { get; }
 
+        private static int R(float f) => (int)MathF.Round(f, MidpointRounding.AwayFromZero);
+
+        // TODO: Move this to a proper test class!
+        [Theory]
+        [InlineData(0.32, 54.5, -3.5, -4.1)]
+        [InlineData(5.3, 536.4, 4.5, 8.1)]
+        public void Vector4_PseudoRound(float x, float y, float z, float w)
+        {
+            var v = new Vector4(x, y, z, w);
+
+            Vector4 actual = v.PseudoRound();
+
+            Assert.Equal(
+                R(v.X),
+                (int)actual.X
+                );
+            Assert.Equal(
+                R(v.Y),
+                (int)actual.Y
+            );
+            Assert.Equal(
+                R(v.Z),
+                (int)actual.Z
+            );
+            Assert.Equal(
+                R(v.W),
+                (int)actual.W
+            );
+        }
+
         [Theory]
         [MemberData(nameof(CommonConversionData))]
         public void ConvertFromYCbCr(int inputBufferLength, int resultBufferLength, int seed)
         {
-            ValidateConversion(
+            ValidateConversion(JpegColorSpace.YCbCr, 3, inputBufferLength, resultBufferLength, seed, ValidateYCbCr);
-                JpegColorSpace.YCbCr,
+        }
-                3,
+
-                inputBufferLength,
+        private static void ValidateYCbCr(JpegColorConverter.ComponentValues values, Span<Vector4> result, int i)
-                resultBufferLength,
-                seed,
-                (values, result, i) =>
         {
             float y = values.Component0[i];
             float cb = values.Component1[i];
@@ -55,7 +83,12 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
             Assert.True(actual.AlmostEquals(expected, Precision));
             Assert.Equal(1, rgba.W);
-                    });
+        }
+
+        [Fact]
+        public void ConvertFromYCbCr_SimdWithAlignedValues()
+        {
+            ValidateConversion(JpegColorConverter.FromYCbCrSimd256.ConvertAligned, 3, 64, 64, 1, ValidateYCbCr);
         }
 
         [Theory]
@@ -207,18 +240,33 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             int inputBufferLength,
             int resultBufferLength,
             int seed,
-            Action<JpegColorConverter.ComponentValues, Span<Vector4>, int> doValidate)
+            Action<JpegColorConverter.ComponentValues, Span<Vector4>, int> validatePixelValue)
         {
-            var converter = JpegColorConverter.GetConverter(colorSpace);
+            ValidateConversion(
+                (v, r) => JpegColorConverter.GetConverter(colorSpace).ConvertToRGBA(v, r),
+                componentCount,
+                inputBufferLength,
+                resultBufferLength,
+                seed,
+                validatePixelValue);
+        }
 
+        private static void ValidateConversion(
+            Action<JpegColorConverter.ComponentValues, Span<Vector4>> doConvert,
+            int componentCount,
+            int inputBufferLength,
+            int resultBufferLength,
+            int seed,
+            Action<JpegColorConverter.ComponentValues, Span<Vector4>, int> validatePixelValue)
+        {
             JpegColorConverter.ComponentValues values = CreateRandomValues(componentCount, inputBufferLength, seed);
             Vector4[] result = new Vector4[resultBufferLength];
 
-            converter.ConvertToRGBA(values, result);
+            doConvert(values, result);
 
             for (int i = 0; i < resultBufferLength; i++)
             {
-                doValidate(values, result, i);
+                validatePixelValue(values, result, i);
             }
         }
     }