Merge pull request #337 from SixLabors/antonfirsov/jpeg-optimization

Accuracy bugfix + further optimizations on JpegImagePostprocessor

src/ImageSharp/Common/Extensions/SimdUtils.cs

@@ -15,9 +15,9 @@ namespace SixLabors.ImageSharp
     internal static class SimdUtils
     {
         /// <summary>
-        /// Indicates AVX2 architecture where both float and integer registers are of size 256 byte.
+        /// Gets a value indicating whether the code is being executed on AVX2 CPU where both float and integer registers are of size 256 byte.
         /// </summary>
-        public static readonly bool IsAvx2CompatibleArchitecture = Vector<float>.Count == 8 && Vector<int>.Count == 8;
+        public static bool IsAvx2CompatibleArchitecture => Vector<float>.Count == 8 && Vector<int>.Count == 8;
 
         internal static void GuardAvx2(string operation)
         {

src/ImageSharp/Common/Tuples/Tuple8.cs

@@ -0,0 +1,98 @@
+using System.Runtime.InteropServices;
+
+namespace SixLabors.ImageSharp.Common.Tuples
+{
+    /// <summary>
+    /// Contains value type tuples of 8 elements.
+    /// TODO: We should T4 this stuff to be DRY
+    /// </summary>
+    internal static class Tuple8
+    {
+        /// <summary>
+        /// Value type tuple of 8 <see cref="uint"/>-s
+        /// </summary>
+        [StructLayout(LayoutKind.Explicit, Size = 8 * sizeof(uint))]
+        public struct OfUInt32
+        {
+            [FieldOffset(0 * sizeof(uint))]
+            public uint V0;
+
+            [FieldOffset(1 * sizeof(uint))]
+            public uint V1;
+
+            [FieldOffset(2 * sizeof(uint))]
+            public uint V2;
+
+            [FieldOffset(3 * sizeof(uint))]
+            public uint V3;
+
+            [FieldOffset(4 * sizeof(uint))]
+            public uint V4;
+
+            [FieldOffset(5 * sizeof(uint))]
+            public uint V5;
+
+            [FieldOffset(6 * sizeof(uint))]
+            public uint V6;
+
+            [FieldOffset(7 * sizeof(uint))]
+            public uint V7;
+
+            public override string ToString()
+            {
+                return $"[{this.V0},{this.V1},{this.V2},{this.V3},{this.V4},{this.V5},{this.V6},{this.V7}]";
+            }
+        }
+
+        /// <summary>
+        /// Value type tuple of 8 <see cref="byte"/>-s
+        /// </summary>
+        [StructLayout(LayoutKind.Explicit, Size = 8)]
+        public struct OfByte
+        {
+            [FieldOffset(0)]
+            public byte V0;
+
+            [FieldOffset(1)]
+            public byte V1;
+
+            [FieldOffset(2)]
+            public byte V2;
+
+            [FieldOffset(3)]
+            public byte V3;
+
+            [FieldOffset(4)]
+            public byte V4;
+
+            [FieldOffset(5)]
+            public byte V5;
+
+            [FieldOffset(6)]
+            public byte V6;
+
+            [FieldOffset(7)]
+            public byte V7;
+
+            public override string ToString()
+            {
+                return $"[{this.V0},{this.V1},{this.V2},{this.V3},{this.V4},{this.V5},{this.V6},{this.V7}]";
+            }
+
+            /// <summary>
+            /// Sets the values of this tuple by casting all elements of the given <see cref="OfUInt32"/> tuple to <see cref="byte"/>.
+            /// </summary>
+            public void LoadFrom(ref OfUInt32 i)
+            {
+                this.V0 = (byte)i.V0;
+                this.V1 = (byte)i.V1;
+                this.V2 = (byte)i.V2;
+                this.V3 = (byte)i.V3;
+                this.V4 = (byte)i.V4;
+                this.V5 = (byte)i.V5;
+                this.V6 = (byte)i.V6;
+                this.V7 = (byte)i.V7;
+            }
+        }
+    }
+}

src/ImageSharp/Common/Tuples/Vector4Pair.cs

@@ -0,0 +1,75 @@
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+
+namespace SixLabors.ImageSharp.Common.Tuples
+{
+    /// <summary>
+    /// Its faster to process multiple Vector4-s together, so let's pair them!
+    /// On AVX2 this pair should be convertible to <see cref="Vector{T}"/> of <see cref="float"/>!
+    /// </summary>
+    [StructLayout(LayoutKind.Sequential)]
+    internal struct Vector4Pair
+    {
+        public Vector4 A;
+
+        public Vector4 B;
+
+        private static readonly Vector4 Scale = new Vector4(1 / 255f);
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void MultiplyInplace(float value)
+        {
+            this.A *= value;
+            this.B *= value;
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void AddInplace(Vector4 value)
+        {
+            this.A += value;
+            this.B += value;
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void AddInplace(ref Vector4Pair other)
+        {
+            this.A += other.A;
+            this.B += other.B;
+        }
+
+        /// <summary>
+        /// Downscale method, specific to Jpeg color conversion. Works only if Vector{float}.Count == 4!
+        /// TODO: Move it somewhere else.
+        /// </summary>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        internal void RoundAndDownscalePreAvx2()
+        {
+            ref Vector<float> a = ref Unsafe.As<Vector4, Vector<float>>(ref this.A);
+            a = a.FastRound();
+
+            ref Vector<float> b = ref Unsafe.As<Vector4, Vector<float>>(ref this.B);
+            b = b.FastRound();
+
+            // Downscale by 1/255
+            this.A *= Scale;
+            this.B *= Scale;
+        }
+
+        /// <summary>
+        /// AVX2-only Downscale method, specific to Jpeg color conversion.
+        /// TODO: Move it somewhere else.
+        /// </summary>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        internal void RoundAndDownscaleAvx2()
+        {
+            ref Vector<float> self = ref Unsafe.As<Vector4Pair, Vector<float>>(ref this);
+            Vector<float> v = self;
+            v = v.FastRound();
+
+            // Downscale by 1/255
+            v *= new Vector<float>(1 / 255f);
+            self = v;
+        }
+    }
+}

src/ImageSharp/Formats/Jpeg/Common/Block8x8.cs

@@ -176,17 +176,12 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
         }
 
         /// <summary>
-        /// Convert into <see cref="Block8x8F"/>
+        /// Convert to <see cref="Block8x8F"/>
         /// </summary>
         public Block8x8F AsFloatBlock()
         {
-            // TODO: Optimize this
             var result = default(Block8x8F);
-            for (int i = 0; i < Size; i++)
-            {
-                result[i] = this[i];
-            }
-
+            result.LoadFrom(ref this);
             return result;
         }
 

src/ImageSharp/Formats/Jpeg/Common/Block8x8F.CopyTo.cs

@@ -0,0 +1,146 @@
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using SixLabors.ImageSharp.Memory;
+
+// ReSharper disable InconsistentNaming
+namespace SixLabors.ImageSharp.Formats.Jpeg.Common
+{
+    internal partial struct Block8x8F
+    {
+        /// <summary>
+        /// Copy block data into the destination color buffer pixel area with the provided horizontal and vertical.
+        /// </summary>
+        public void CopyTo(BufferArea<float> area, int horizontalScale, int verticalScale)
+        {
+            if (horizontalScale == 1 && verticalScale == 1)
+            {
+                this.CopyTo(area);
+                return;
+            }
+            else if (horizontalScale == 2 && verticalScale == 2)
+            {
+                this.CopyTo2x2(area);
+                return;
+            }
+
+            // TODO: Optimize: implement all the cases with loopless special code! (T4?)
+            for (int y = 0; y < 8; y++)
+            {
+                int yy = y * verticalScale;
+                int y8 = y * 8;
+
+                for (int x = 0; x < 8; x++)
+                {
+                    int xx = x * horizontalScale;
+
+                    float value = this[y8 + x];
+
+                    for (int i = 0; i < verticalScale; i++)
+                    {
+                        for (int j = 0; j < horizontalScale; j++)
+                        {
+                            area[xx + j, yy + i] = value;
+                        }
+                    }
+                }
+            }
+        }
+
+        // [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void CopyTo(BufferArea<float> area)
+        {
+            ref byte selfBase = ref Unsafe.As<Block8x8F, byte>(ref this);
+            ref byte destBase = ref Unsafe.As<float, byte>(ref area.GetReferenceToOrigo());
+            int destStride = area.Stride * sizeof(float);
+
+            CopyRowImpl(ref selfBase, ref destBase, destStride, 0);
+            CopyRowImpl(ref selfBase, ref destBase, destStride, 1);
+            CopyRowImpl(ref selfBase, ref destBase, destStride, 2);
+            CopyRowImpl(ref selfBase, ref destBase, destStride, 3);
+            CopyRowImpl(ref selfBase, ref destBase, destStride, 4);
+            CopyRowImpl(ref selfBase, ref destBase, destStride, 5);
+            CopyRowImpl(ref selfBase, ref destBase, destStride, 6);
+            CopyRowImpl(ref selfBase, ref destBase, destStride, 7);
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static void CopyRowImpl(ref byte selfBase, ref byte destBase, int destStride, int row)
+        {
+            ref byte s = ref Unsafe.Add(ref selfBase, row * 8 * sizeof(float));
+            ref byte d = ref Unsafe.Add(ref destBase, row * destStride);
+            Unsafe.CopyBlock(ref d, ref s, 8 * sizeof(float));
+        }
+
+        private void CopyTo2x2(BufferArea<float> area)
+        {
+            ref float destBase = ref area.GetReferenceToOrigo();
+            int destStride = area.Stride;
+
+            this.WidenCopyImpl2x2(ref destBase, 0, destStride);
+            this.WidenCopyImpl2x2(ref destBase, 1, destStride);
+            this.WidenCopyImpl2x2(ref destBase, 2, destStride);
+            this.WidenCopyImpl2x2(ref destBase, 3, destStride);
+            this.WidenCopyImpl2x2(ref destBase, 4, destStride);
+            this.WidenCopyImpl2x2(ref destBase, 5, destStride);
+            this.WidenCopyImpl2x2(ref destBase, 6, destStride);
+            this.WidenCopyImpl2x2(ref destBase, 7, destStride);
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private void WidenCopyImpl2x2(ref float destBase, int row, int destStride)
+        {
+            ref Vector4 selfLeft = ref Unsafe.Add(ref this.V0L, 2 * row);
+            ref Vector4 selfRight = ref Unsafe.Add(ref selfLeft, 1);
+            ref float destLocalOrigo = ref Unsafe.Add(ref destBase, row * 2 * destStride);
+
+            Unsafe.Add(ref destLocalOrigo, 0) = selfLeft.X;
+            Unsafe.Add(ref destLocalOrigo, 1) = selfLeft.X;
+            Unsafe.Add(ref destLocalOrigo, 2) = selfLeft.Y;
+            Unsafe.Add(ref destLocalOrigo, 3) = selfLeft.Y;
+            Unsafe.Add(ref destLocalOrigo, 4) = selfLeft.Z;
+            Unsafe.Add(ref destLocalOrigo, 5) = selfLeft.Z;
+            Unsafe.Add(ref destLocalOrigo, 6) = selfLeft.W;
+            Unsafe.Add(ref destLocalOrigo, 7) = selfLeft.W;
+
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, 8), 0) = selfRight.X;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, 8), 1) = selfRight.X;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, 8), 2) = selfRight.Y;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, 8), 3) = selfRight.Y;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, 8), 4) = selfRight.Z;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, 8), 5) = selfRight.Z;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, 8), 6) = selfRight.W;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, 8), 7) = selfRight.W;
+
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride), 0) = selfLeft.X;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride), 1) = selfLeft.X;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride), 2) = selfLeft.Y;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride), 3) = selfLeft.Y;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride), 4) = selfLeft.Z;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride), 5) = selfLeft.Z;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride), 6) = selfLeft.W;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride), 7) = selfLeft.W;
+
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride + 8), 0) = selfRight.X;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride + 8), 1) = selfRight.X;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride + 8), 2) = selfRight.Y;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride + 8), 3) = selfRight.Y;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride + 8), 4) = selfRight.Z;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride + 8), 5) = selfRight.Z;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride + 8), 6) = selfRight.W;
+            Unsafe.Add(ref Unsafe.Add(ref destLocalOrigo, destStride + 8), 7) = selfRight.W;
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static void WidenCopyImpl(ref Vector4 s, ref float destBase)
+        {
+            Unsafe.Add(ref destBase, 0) = s.X;
+            Unsafe.Add(ref destBase, 1) = s.X;
+            Unsafe.Add(ref destBase, 2) = s.Y;
+            Unsafe.Add(ref destBase, 3) = s.Y;
+            Unsafe.Add(ref destBase, 4) = s.Z;
+            Unsafe.Add(ref destBase, 5) = s.Z;
+            Unsafe.Add(ref destBase, 6) = s.W;
+            Unsafe.Add(ref destBase, 7) = s.W;
+        }
+    }
+}

src/ImageSharp/Formats/Jpeg/Common/Block8x8F.Generated.cs

@@ -96,51 +96,139 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
 		/// <summary>
         /// Level shift by +128, clip to [0, 255]
         /// </summary>
-        /// <param name="d">The destination block</param>
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        internal void NormalizeColorsInto(ref Block8x8F d)
+        public void NormalizeColorsInplace()
         {
-            d.V0L = Vector4.Clamp(V0L + COff4, CMin4, CMax4);
-            d.V0R = Vector4.Clamp(V0R + COff4, CMin4, CMax4);
-            d.V1L = Vector4.Clamp(V1L + COff4, CMin4, CMax4);
-            d.V1R = Vector4.Clamp(V1R + COff4, CMin4, CMax4);
-            d.V2L = Vector4.Clamp(V2L + COff4, CMin4, CMax4);
-            d.V2R = Vector4.Clamp(V2R + COff4, CMin4, CMax4);
-            d.V3L = Vector4.Clamp(V3L + COff4, CMin4, CMax4);
-            d.V3R = Vector4.Clamp(V3R + COff4, CMin4, CMax4);
-            d.V4L = Vector4.Clamp(V4L + COff4, CMin4, CMax4);
-            d.V4R = Vector4.Clamp(V4R + COff4, CMin4, CMax4);
-            d.V5L = Vector4.Clamp(V5L + COff4, CMin4, CMax4);
-            d.V5R = Vector4.Clamp(V5R + COff4, CMin4, CMax4);
-            d.V6L = Vector4.Clamp(V6L + COff4, CMin4, CMax4);
-            d.V6R = Vector4.Clamp(V6R + COff4, CMin4, CMax4);
-            d.V7L = Vector4.Clamp(V7L + COff4, CMin4, CMax4);
-            d.V7R = Vector4.Clamp(V7R + COff4, CMin4, CMax4);
+            this.V0L = Vector4.Clamp(this.V0L + COff4, CMin4, CMax4);
+            this.V0R = Vector4.Clamp(this.V0R + COff4, CMin4, CMax4);
+            this.V1L = Vector4.Clamp(this.V1L + COff4, CMin4, CMax4);
+            this.V1R = Vector4.Clamp(this.V1R + COff4, CMin4, CMax4);
+            this.V2L = Vector4.Clamp(this.V2L + COff4, CMin4, CMax4);
+            this.V2R = Vector4.Clamp(this.V2R + COff4, CMin4, CMax4);
+            this.V3L = Vector4.Clamp(this.V3L + COff4, CMin4, CMax4);
+            this.V3R = Vector4.Clamp(this.V3R + COff4, CMin4, CMax4);
+            this.V4L = Vector4.Clamp(this.V4L + COff4, CMin4, CMax4);
+            this.V4R = Vector4.Clamp(this.V4R + COff4, CMin4, CMax4);
+            this.V5L = Vector4.Clamp(this.V5L + COff4, CMin4, CMax4);
+            this.V5R = Vector4.Clamp(this.V5R + COff4, CMin4, CMax4);
+            this.V6L = Vector4.Clamp(this.V6L + COff4, CMin4, CMax4);
+            this.V6R = Vector4.Clamp(this.V6R + COff4, CMin4, CMax4);
+            this.V7L = Vector4.Clamp(this.V7L + COff4, CMin4, CMax4);
+            this.V7R = Vector4.Clamp(this.V7R + COff4, CMin4, CMax4);
         }
 
-
         /// <summary>
-        /// Level shift by +128, clip to [0, 255]
+        /// AVX2-only variant for executing <see cref="NormalizeColorsInplace"/> and <see cref="RoundInplace"/> in one step.
         /// </summary>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        internal void NormalizeColorsInplace()
+        public void NormalizeColorsAndRoundInplaceAvx2()
+        {
+            Vector<float> off = new Vector<float>(128f);
+            Vector<float> max = new Vector<float>(255F);
+            
+            ref Vector<float> row0 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V0L);
+            row0 = NormalizeAndRound(row0, off, max);
+                
+            ref Vector<float> row1 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V1L);
+            row1 = NormalizeAndRound(row1, off, max);
+                
+            ref Vector<float> row2 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V2L);
+            row2 = NormalizeAndRound(row2, off, max);
+                
+            ref Vector<float> row3 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V3L);
+            row3 = NormalizeAndRound(row3, off, max);
+                
+            ref Vector<float> row4 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V4L);
+            row4 = NormalizeAndRound(row4, off, max);
+                
+            ref Vector<float> row5 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V5L);
+            row5 = NormalizeAndRound(row5, off, max);
+                
+            ref Vector<float> row6 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V6L);
+            row6 = NormalizeAndRound(row6, off, max);
+                
+            ref Vector<float> row7 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V7L);
+            row7 = NormalizeAndRound(row7, off, max);
+                
+        }
+
+        /// <summary>
+        /// Fill the block from 'source' doing short -> float conversion.
+        /// </summary>
+        public void LoadFrom(ref Block8x8 source)
         {
-            this.V0L = Vector4.Clamp(V0L + COff4, CMin4, CMax4);
-            this.V0R = Vector4.Clamp(V0R + COff4, CMin4, CMax4);
-            this.V1L = Vector4.Clamp(V1L + COff4, CMin4, CMax4);
-            this.V1R = Vector4.Clamp(V1R + COff4, CMin4, CMax4);
-            this.V2L = Vector4.Clamp(V2L + COff4, CMin4, CMax4);
-            this.V2R = Vector4.Clamp(V2R + COff4, CMin4, CMax4);
-            this.V3L = Vector4.Clamp(V3L + COff4, CMin4, CMax4);
-            this.V3R = Vector4.Clamp(V3R + COff4, CMin4, CMax4);
-            this.V4L = Vector4.Clamp(V4L + COff4, CMin4, CMax4);
-            this.V4R = Vector4.Clamp(V4R + COff4, CMin4, CMax4);
-            this.V5L = Vector4.Clamp(V5L + COff4, CMin4, CMax4);
-            this.V5R = Vector4.Clamp(V5R + COff4, CMin4, CMax4);
-            this.V6L = Vector4.Clamp(V6L + COff4, CMin4, CMax4);
-            this.V6R = Vector4.Clamp(V6R + COff4, CMin4, CMax4);
-            this.V7L = Vector4.Clamp(V7L + COff4, CMin4, CMax4);
-            this.V7R = Vector4.Clamp(V7R + COff4, CMin4, CMax4);
+            ref short selfRef = ref Unsafe.As<Block8x8, short>(ref source);
+
+            this.V0L.X =  Unsafe.Add(ref selfRef, 0);
+            this.V0L.Y =  Unsafe.Add(ref selfRef, 1);
+            this.V0L.Z =  Unsafe.Add(ref selfRef, 2);
+            this.V0L.W =  Unsafe.Add(ref selfRef, 3);
+            this.V0R.X =  Unsafe.Add(ref selfRef, 4);
+            this.V0R.Y =  Unsafe.Add(ref selfRef, 5);
+            this.V0R.Z =  Unsafe.Add(ref selfRef, 6);
+            this.V0R.W =  Unsafe.Add(ref selfRef, 7);
+
+            this.V1L.X =  Unsafe.Add(ref selfRef, 8);
+            this.V1L.Y =  Unsafe.Add(ref selfRef, 9);
+            this.V1L.Z =  Unsafe.Add(ref selfRef, 10);
+            this.V1L.W =  Unsafe.Add(ref selfRef, 11);
+            this.V1R.X =  Unsafe.Add(ref selfRef, 12);
+            this.V1R.Y =  Unsafe.Add(ref selfRef, 13);
+            this.V1R.Z =  Unsafe.Add(ref selfRef, 14);
+            this.V1R.W =  Unsafe.Add(ref selfRef, 15);
+
+            this.V2L.X =  Unsafe.Add(ref selfRef, 16);
+            this.V2L.Y =  Unsafe.Add(ref selfRef, 17);
+            this.V2L.Z =  Unsafe.Add(ref selfRef, 18);
+            this.V2L.W =  Unsafe.Add(ref selfRef, 19);
+            this.V2R.X =  Unsafe.Add(ref selfRef, 20);
+            this.V2R.Y =  Unsafe.Add(ref selfRef, 21);
+            this.V2R.Z =  Unsafe.Add(ref selfRef, 22);
+            this.V2R.W =  Unsafe.Add(ref selfRef, 23);
+
+            this.V3L.X =  Unsafe.Add(ref selfRef, 24);
+            this.V3L.Y =  Unsafe.Add(ref selfRef, 25);
+            this.V3L.Z =  Unsafe.Add(ref selfRef, 26);
+            this.V3L.W =  Unsafe.Add(ref selfRef, 27);
+            this.V3R.X =  Unsafe.Add(ref selfRef, 28);
+            this.V3R.Y =  Unsafe.Add(ref selfRef, 29);
+            this.V3R.Z =  Unsafe.Add(ref selfRef, 30);
+            this.V3R.W =  Unsafe.Add(ref selfRef, 31);
+
+            this.V4L.X =  Unsafe.Add(ref selfRef, 32);
+            this.V4L.Y =  Unsafe.Add(ref selfRef, 33);
+            this.V4L.Z =  Unsafe.Add(ref selfRef, 34);
+            this.V4L.W =  Unsafe.Add(ref selfRef, 35);
+            this.V4R.X =  Unsafe.Add(ref selfRef, 36);
+            this.V4R.Y =  Unsafe.Add(ref selfRef, 37);
+            this.V4R.Z =  Unsafe.Add(ref selfRef, 38);
+            this.V4R.W =  Unsafe.Add(ref selfRef, 39);
+
+            this.V5L.X =  Unsafe.Add(ref selfRef, 40);
+            this.V5L.Y =  Unsafe.Add(ref selfRef, 41);
+            this.V5L.Z =  Unsafe.Add(ref selfRef, 42);
+            this.V5L.W =  Unsafe.Add(ref selfRef, 43);
+            this.V5R.X =  Unsafe.Add(ref selfRef, 44);
+            this.V5R.Y =  Unsafe.Add(ref selfRef, 45);
+            this.V5R.Z =  Unsafe.Add(ref selfRef, 46);
+            this.V5R.W =  Unsafe.Add(ref selfRef, 47);
+
+            this.V6L.X =  Unsafe.Add(ref selfRef, 48);
+            this.V6L.Y =  Unsafe.Add(ref selfRef, 49);
+            this.V6L.Z =  Unsafe.Add(ref selfRef, 50);
+            this.V6L.W =  Unsafe.Add(ref selfRef, 51);
+            this.V6R.X =  Unsafe.Add(ref selfRef, 52);
+            this.V6R.Y =  Unsafe.Add(ref selfRef, 53);
+            this.V6R.Z =  Unsafe.Add(ref selfRef, 54);
+            this.V6R.W =  Unsafe.Add(ref selfRef, 55);
+
+            this.V7L.X =  Unsafe.Add(ref selfRef, 56);
+            this.V7L.Y =  Unsafe.Add(ref selfRef, 57);
+            this.V7L.Z =  Unsafe.Add(ref selfRef, 58);
+            this.V7L.W =  Unsafe.Add(ref selfRef, 59);
+            this.V7R.X =  Unsafe.Add(ref selfRef, 60);
+            this.V7R.Y =  Unsafe.Add(ref selfRef, 61);
+            this.V7R.Z =  Unsafe.Add(ref selfRef, 62);
+            this.V7R.W =  Unsafe.Add(ref selfRef, 63);
         }
-    }
+	}
 }

src/ImageSharp/Formats/Jpeg/Common/Block8x8F.Generated.tt

@@ -61,9 +61,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
 		/// <summary>
         /// Level shift by +128, clip to [0, 255]
         /// </summary>
-        /// <param name="d">The destination block</param>
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        internal void TransformByteConvetibleColorValuesInto(ref Block8x8F d)
+        public void NormalizeColorsInplace()
         {
             <#
 
@@ -74,7 +72,61 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
                 for (int j = 0; j < 2; j++)
                 {
 					char side = j == 0 ? 'L' : 'R';
-					Write($"d.V{i}{side} = Vector4.Clamp(V{i}{side} + COff4, CMin4, CMax4);\r\n");
+					Write($"this.V{i}{side} = Vector4.Clamp(this.V{i}{side} + COff4, CMin4, CMax4);\r\n");
+                }
+            }
+			PopIndent();
+			#>
+        }
+
+        /// <summary>
+        /// AVX2-only variant for executing <see cref="NormalizeColorsInplace"/> and <see cref="RoundInplace"/> in one step.
+        /// </summary>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void NormalizeColorsAndRoundInplaceAvx2()
+        {
+            Vector<float> off = new Vector<float>(128f);
+            Vector<float> max = new Vector<float>(255F);
+            <#
+
+            for (int i = 0; i < 8; i++)
+            {
+                #>
+
+            ref Vector<float> row<#=i#> = ref Unsafe.As<Vector4, Vector<float>>(ref this.V<#=i#>L);
+            row<#=i#> = NormalizeAndRound(row<#=i#>, off, max);
+                <#
+            }
+			#>
+
+        }
+
+        /// <summary>
+        /// Fill the block from 'source' doing short -> float conversion.
+        /// </summary>
+        public void LoadFrom(ref Block8x8 source)
+        {
+            ref short selfRef = ref Unsafe.As<Block8x8, short>(ref source);
+
+            <#
+			PushIndent("            ");
+            for (int j = 0; j < 8; j++)
+            {
+                for (int i = 0; i < 8; i++)
+                {
+                    char destCoord = coordz[i % 4];
+                    char destSide = (i / 4) % 2 == 0 ? 'L' : 'R';
+                                    
+					if(j > 0 && i == 0){
+					WriteLine("");
+					}
+
+                    char srcCoord = coordz[j % 4];
+                    char srcSide = (j / 4) % 2 == 0 ? 'L' : 'R';
+
+                    string expression = $"this.V{j}{destSide}.{destCoord} =  Unsafe.Add(ref selfRef, {j*8+i});\r\n";
+					Write(expression);
+                
                 }
             }
 			PopIndent();

src/ImageSharp/Formats/Jpeg/Common/Block8x8F.cs

@@ -7,7 +7,6 @@ using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 using System.Text;
-using SixLabors.ImageSharp.Memory;
 
 // ReSharper disable InconsistentNaming
 namespace SixLabors.ImageSharp.Formats.Jpeg.Common
@@ -17,16 +16,6 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
     /// </summary>
     internal partial struct Block8x8F
     {
-        // Most of the static methods of this struct are instance methods by actual semantics: they use Block8x8F* as their first parameter.
-        // Example: GetScalarAt() and SetScalarAt() are really just other (optimized) versions of the indexer.
-        // It's much cleaner, easier and safer to work with the code, if the methods with same semantics are next to each other.
-#pragma warning disable SA1204 // StaticElementsMustAppearBeforeInstanceElements
-
-        /// <summary>
-        /// Vector count
-        /// </summary>
-        public const int VectorCount = 16;
-
         /// <summary>
         /// A number of scalar coefficients in a <see cref="Block8x8F"/>
         /// </summary>
@@ -157,36 +146,6 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
             return result;
         }
 
-        /// <summary>
-        /// Pointer-based "Indexer" (getter part)
-        /// </summary>
-        /// <param name="blockPtr">Block pointer</param>
-        /// <param name="idx">Index</param>
-        /// <returns>The scaleVec value at the specified index</returns>
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static unsafe float GetScalarAt(Block8x8F* blockPtr, int idx)
-        {
-            GuardBlockIndex(idx);
-
-            float* fp = (float*)blockPtr;
-            return fp[idx];
-        }
-
-        /// <summary>
-        /// Pointer-based "Indexer" (setter part)
-        /// </summary>
-        /// <param name="blockPtr">Block pointer</param>
-        /// <param name="idx">Index</param>
-        /// <param name="value">Value</param>
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public static unsafe void SetScalarAt(Block8x8F* blockPtr, int idx, float value)
-        {
-            GuardBlockIndex(idx);
-
-            float* fp = (float*)blockPtr;
-            fp[idx] = value;
-        }
-
         /// <summary>
         /// Fill the block with defaults (zeroes)
         /// </summary>
@@ -242,7 +201,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
         /// </summary>
         /// <param name="dest">Destination</param>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public unsafe void CopyTo(Span<float> dest)
+        public void CopyTo(Span<float> dest)
         {
             ref byte d = ref Unsafe.As<float, byte>(ref dest.DangerousGetPinnableReference());
             ref byte s = ref Unsafe.As<Block8x8F, byte>(ref this);
@@ -306,109 +265,6 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
             }
         }
 
-        // [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public void CopyTo(BufferArea<float> area)
-        {
-            ref byte selfBase = ref Unsafe.As<Block8x8F, byte>(ref this);
-            ref byte destBase = ref Unsafe.As<float, byte>(ref area.GetReferenceToOrigo());
-            int destStride = area.Stride * sizeof(float);
-
-            CopyRowImpl(ref selfBase, ref destBase, destStride, 0);
-            CopyRowImpl(ref selfBase, ref destBase, destStride, 1);
-            CopyRowImpl(ref selfBase, ref destBase, destStride, 2);
-            CopyRowImpl(ref selfBase, ref destBase, destStride, 3);
-            CopyRowImpl(ref selfBase, ref destBase, destStride, 4);
-            CopyRowImpl(ref selfBase, ref destBase, destStride, 5);
-            CopyRowImpl(ref selfBase, ref destBase, destStride, 6);
-            CopyRowImpl(ref selfBase, ref destBase, destStride, 7);
-        }
-
-        public void CopyTo(BufferArea<float> area, int horizontalScale, int verticalScale)
-        {
-            if (horizontalScale == 1 && verticalScale == 1)
-            {
-                this.CopyTo(area);
-                return;
-            }
-            else if (horizontalScale == 2 && verticalScale == 2)
-            {
-                this.CopyTo2x2(area);
-                return;
-            }
-
-            // TODO: Optimize: implement all the cases with loopless special code! (T4?)
-            for (int y = 0; y < 8; y++)
-            {
-                int yy = y * verticalScale;
-                int y8 = y * 8;
-
-                for (int x = 0; x < 8; x++)
-                {
-                    int xx = x * horizontalScale;
-
-                    float value = this[y8 + x];
-
-                    for (int i = 0; i < verticalScale; i++)
-                    {
-                        for (int j = 0; j < horizontalScale; j++)
-                        {
-                            area[xx + j, yy + i] = value;
-                        }
-                    }
-                }
-            }
-        }
-
-        private void CopyTo2x2(BufferArea<float> area)
-        {
-            ref float destBase = ref area.GetReferenceToOrigo();
-            int destStride = area.Stride;
-
-            this.CopyRow2x2Impl(ref destBase, 0, destStride);
-            this.CopyRow2x2Impl(ref destBase, 1, destStride);
-            this.CopyRow2x2Impl(ref destBase, 2, destStride);
-            this.CopyRow2x2Impl(ref destBase, 3, destStride);
-            this.CopyRow2x2Impl(ref destBase, 4, destStride);
-            this.CopyRow2x2Impl(ref destBase, 5, destStride);
-            this.CopyRow2x2Impl(ref destBase, 6, destStride);
-            this.CopyRow2x2Impl(ref destBase, 7, destStride);
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        private static void CopyRowImpl(ref byte selfBase, ref byte destBase, int destStride, int row)
-        {
-            ref byte s = ref Unsafe.Add(ref selfBase, row * 8 * sizeof(float));
-            ref byte d = ref Unsafe.Add(ref destBase, row * destStride);
-            Unsafe.CopyBlock(ref d, ref s, 8 * sizeof(float));
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        private void CopyRow2x2Impl(ref float destBase, int row, int destStride)
-        {
-            ref Vector4 selfLeft = ref Unsafe.Add(ref this.V0L, 2 * row);
-            ref Vector4 selfRight = ref Unsafe.Add(ref selfLeft, 1);
-            ref float destLocalOrigo = ref Unsafe.Add(ref destBase, row * 2 * destStride);
-
-            Stride2VectorCopyImpl(ref selfLeft, ref destLocalOrigo);
-            Stride2VectorCopyImpl(ref selfRight, ref Unsafe.Add(ref destLocalOrigo, 8));
-
-            Stride2VectorCopyImpl(ref selfLeft, ref Unsafe.Add(ref destLocalOrigo, destStride));
-            Stride2VectorCopyImpl(ref selfRight, ref Unsafe.Add(ref destLocalOrigo, destStride + 8));
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        private static void Stride2VectorCopyImpl(ref Vector4 s, ref float destBase)
-        {
-            Unsafe.Add(ref destBase, 0) = s.X;
-            Unsafe.Add(ref destBase, 1) = s.X;
-            Unsafe.Add(ref destBase, 2) = s.Y;
-            Unsafe.Add(ref destBase, 3) = s.Y;
-            Unsafe.Add(ref destBase, 4) = s.Z;
-            Unsafe.Add(ref destBase, 5) = s.Z;
-            Unsafe.Add(ref destBase, 6) = s.W;
-            Unsafe.Add(ref destBase, 7) = s.W;
-        }
-
         public float[] ToArray()
         {
             float[] result = new float[Size];
@@ -419,26 +275,50 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
         /// <summary>
         /// Multiply all elements of the block.
         /// </summary>
-        /// <param name="scaleVec">Vector to multiply by</param>
+        /// <param name="value">The value to multiply by</param>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public void MultiplyAllInplace(float scaleVec)
-        {
-            this.V0L *= scaleVec;
-            this.V0R *= scaleVec;
-            this.V1L *= scaleVec;
-            this.V1R *= scaleVec;
-            this.V2L *= scaleVec;
-            this.V2R *= scaleVec;
-            this.V3L *= scaleVec;
-            this.V3R *= scaleVec;
-            this.V4L *= scaleVec;
-            this.V4R *= scaleVec;
-            this.V5L *= scaleVec;
-            this.V5R *= scaleVec;
-            this.V6L *= scaleVec;
-            this.V6R *= scaleVec;
-            this.V7L *= scaleVec;
-            this.V7R *= scaleVec;
+        public void MultiplyInplace(float value)
+        {
+            this.V0L *= value;
+            this.V0R *= value;
+            this.V1L *= value;
+            this.V1R *= value;
+            this.V2L *= value;
+            this.V2R *= value;
+            this.V3L *= value;
+            this.V3R *= value;
+            this.V4L *= value;
+            this.V4R *= value;
+            this.V5L *= value;
+            this.V5R *= value;
+            this.V6L *= value;
+            this.V6R *= value;
+            this.V7L *= value;
+            this.V7R *= value;
+        }
+
+        /// <summary>
+        /// Multiply all elements of the block by the corresponding elements of 'other'
+        /// </summary>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void MultiplyInplace(ref Block8x8F other)
+        {
+            this.V0L *= other.V0L;
+            this.V0R *= other.V0R;
+            this.V1L *= other.V1L;
+            this.V1R *= other.V1R;
+            this.V2L *= other.V2L;
+            this.V2R *= other.V2R;
+            this.V3L *= other.V3L;
+            this.V3R *= other.V3R;
+            this.V4L *= other.V4L;
+            this.V4R *= other.V4R;
+            this.V5L *= other.V5L;
+            this.V5R *= other.V5R;
+            this.V6L *= other.V6L;
+            this.V6R *= other.V6R;
+            this.V7L *= other.V7L;
+            this.V7R *= other.V7R;
         }
 
         /// <summary>
@@ -472,56 +352,32 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
         /// <param name="blockPtr">Block pointer</param>
         /// <param name="qtPtr">Qt pointer</param>
         /// <param name="unzigPtr">Unzig pointer</param>
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        // [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static unsafe void DequantizeBlock(Block8x8F* blockPtr, Block8x8F* qtPtr, int* unzigPtr)
         {
             float* b = (float*)blockPtr;
             float* qtp = (float*)qtPtr;
-            for (int zig = 0; zig < Size; zig++)
+            for (int qtIndex = 0; qtIndex < Size; qtIndex++)
             {
-                float* unzigPos = b + unzigPtr[zig];
+                int blockIndex = unzigPtr[qtIndex];
+                float* unzigPos = b + blockIndex;
+
                 float val = *unzigPos;
-                val *= qtp[zig];
+                val *= qtp[qtIndex];
                 *unzigPos = val;
             }
         }
 
         /// <summary>
-        /// Level shift by +128, clip to [0, 255], and write to buffer.
-        /// </summary>
-        /// <param name="destinationBuffer">Color buffer</param>
-        /// <param name="stride">Stride offset</param>
-        /// <param name="tempBlockPtr">Temp Block pointer</param>
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public unsafe void CopyColorsTo(Span<byte> destinationBuffer, int stride, Block8x8F* tempBlockPtr)
-        {
-            this.NormalizeColorsInto(ref *tempBlockPtr);
-            ref byte d = ref destinationBuffer.DangerousGetPinnableReference();
-            float* src = (float*)tempBlockPtr;
-            for (int i = 0; i < 8; i++)
-            {
-                ref byte dRow = ref Unsafe.Add(ref d, i * stride);
-                Unsafe.Add(ref dRow, 0) = (byte)src[0];
-                Unsafe.Add(ref dRow, 1) = (byte)src[1];
-                Unsafe.Add(ref dRow, 2) = (byte)src[2];
-                Unsafe.Add(ref dRow, 3) = (byte)src[3];
-                Unsafe.Add(ref dRow, 4) = (byte)src[4];
-                Unsafe.Add(ref dRow, 5) = (byte)src[5];
-                Unsafe.Add(ref dRow, 6) = (byte)src[6];
-                Unsafe.Add(ref dRow, 7) = (byte)src[7];
-                src += 8;
-            }
-        }
-
-        /// <summary>
+        /// Quantize 'block' into 'dest' using the 'qt' quantization table:
         /// Unzig the elements of block into dest, while dividing them by elements of qt and "pre-rounding" the values.
         /// To finish the rounding it's enough to (int)-cast these values.
         /// </summary>
         /// <param name="block">Source block</param>
         /// <param name="dest">Destination block</param>
         /// <param name="qt">The quantization table</param>
-        /// <param name="unzigPtr">Pointer to elements of <see cref="UnzigData"/></param>
-        public static unsafe void UnzigDivRound(
+        /// <param name="unzigPtr">Pointer to elements of <see cref="ZigZag"/></param>
+        public static unsafe void Quantize(
             Block8x8F* block,
             Block8x8F* dest,
             Block8x8F* qt,
@@ -610,34 +466,26 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
             return result;
         }
 
-        public void RoundInplace()
+        /// <summary>
+        /// Level shift by +128, clip to [0..255], and round all the values in the block.
+        /// </summary>
+        public void NormalizeColorsAndRoundInplace()
         {
-            if (Vector<float>.Count == 8 && Vector<int>.Count == 8)
+            if (SimdUtils.IsAvx2CompatibleArchitecture)
             {
-                ref Vector<float> row0 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V0L);
-                row0 = row0.FastRound();
-                ref Vector<float> row1 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V1L);
-                row1 = row1.FastRound();
-                ref Vector<float> row2 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V2L);
-                row2 = row2.FastRound();
-                ref Vector<float> row3 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V3L);
-                row3 = row3.FastRound();
-                ref Vector<float> row4 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V4L);
-                row4 = row4.FastRound();
-                ref Vector<float> row5 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V5L);
-                row5 = row5.FastRound();
-                ref Vector<float> row6 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V6L);
-                row6 = row6.FastRound();
-                ref Vector<float> row7 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V7L);
-                row7 = row7.FastRound();
+                this.NormalizeColorsAndRoundInplaceAvx2();
             }
             else
             {
-                this.RoundInplaceSlow();
+                this.NormalizeColorsInplace();
+                this.RoundInplace();
             }
         }
 
-        private void RoundInplaceSlow()
+        /// <summary>
+        /// Rounds all values in the block.
+        /// </summary>
+        public void RoundInplace()
         {
             for (int i = 0; i < Size; i++)
             {
@@ -663,6 +511,15 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
             return bld.ToString();
         }
 
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static Vector<float> NormalizeAndRound(Vector<float> row, Vector<float> off, Vector<float> max)
+        {
+            row += off;
+            row = Vector.Max(row, Vector<float>.Zero);
+            row = Vector.Min(row, max);
+            return row.FastRound();
+        }
+
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         private static Vector4 DivideRound(Vector4 dividend, Vector4 divisor)
         {
@@ -679,10 +536,5 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
             DebugGuard.MustBeLessThan(idx, Size, nameof(idx));
             DebugGuard.MustBeGreaterThanOrEqualTo(idx, 0, nameof(idx));
         }
-
-        [StructLayout(LayoutKind.Explicit, Size = 8 * sizeof(float))]
-        private struct Row
-        {
-        }
     }
 }

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

@@ -11,145 +11,72 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
     /// Encapsulates the implementation of processing "raw" <see cref="Buffer{T}"/>-s into Jpeg image channels.
     /// </summary>
     [StructLayout(LayoutKind.Sequential)]
-    internal unsafe struct JpegBlockPostProcessor
+    internal struct JpegBlockPostProcessor
     {
         /// <summary>
-        /// The <see cref="ComputationData"/>
+        /// Source block
         /// </summary>
-        private ComputationData data;
+        public Block8x8F SourceBlock;
 
         /// <summary>
-        /// Pointers to elements of <see cref="data"/>
+        /// Temporal block 1 to store intermediate and/or final computation results
         /// </summary>
-        private DataPointers pointers;
+        public Block8x8F WorkspaceBlock1;
 
         /// <summary>
-        /// Initialize the <see cref="JpegBlockPostProcessor"/> instance on the stack.
+        /// Temporal block 2 to store intermediate and/or final computation results
         /// </summary>
-        /// <param name="postProcessor">The <see cref="JpegBlockPostProcessor"/> instance</param>
-        public static void Init(JpegBlockPostProcessor* postProcessor)
-        {
-            postProcessor->data = ComputationData.Create();
-            postProcessor->pointers = new DataPointers(&postProcessor->data);
-        }
-
-        public void QuantizeAndTransform(IRawJpegData decoder, IJpegComponent component, ref Block8x8 sourceBlock)
-        {
-            this.data.SourceBlock = sourceBlock.AsFloatBlock();
-            int qtIndex = component.QuantizationTableIndex;
-            this.data.QuantiazationTable = decoder.QuantizationTables[qtIndex];
-
-            Block8x8F* b = this.pointers.SourceBlock;
-
-            Block8x8F.DequantizeBlock(b, this.pointers.QuantiazationTable, this.pointers.Unzig);
-
-            FastFloatingPointDCT.TransformIDCT(ref *b, ref this.data.WorkspaceBlock1, ref this.data.WorkspaceBlock2);
-        }
+        public Block8x8F WorkspaceBlock2;
 
-        public void ProcessBlockColorsInto(
-            IRawJpegData decoder,
-            IJpegComponent component,
-            ref Block8x8 sourceBlock,
-            BufferArea<float> destArea)
-        {
-            this.QuantizeAndTransform(decoder, component, ref sourceBlock);
-
-            this.data.WorkspaceBlock1.NormalizeColorsInplace();
-
-            // To conform better to libjpeg we actually NEED TO loose precision here.
-            // This is because they store blocks as Int16 between all the operations.
-            // Unfortunately, we need to emulate this to be "more accurate" :(
-            this.data.WorkspaceBlock1.RoundInplace();
+        /// <summary>
+        /// The quantization table as <see cref="Block8x8F"/>
+        /// </summary>
+        public Block8x8F DequantiazationTable;
 
-            Size divs = component.SubSamplingDivisors;
-            this.data.WorkspaceBlock1.CopyTo(destArea, divs.Width, divs.Height);
-        }
+        /// <summary>
+        /// Defines the horizontal and vertical scale we need to apply to the 8x8 sized block.
+        /// </summary>
+        private Size subSamplingDivisors;
 
         /// <summary>
-        /// Holds the "large" data blocks needed for computations.
+        /// Initializes a new instance of the <see cref="JpegBlockPostProcessor"/> struct.
         /// </summary>
-        [StructLayout(LayoutKind.Sequential)]
-        public struct ComputationData
+        public JpegBlockPostProcessor(IRawJpegData decoder, IJpegComponent component)
         {
-            /// <summary>
-            /// Source block
-            /// </summary>
-            public Block8x8F SourceBlock;
-
-            /// <summary>
-            /// Temporal block 1 to store intermediate and/or final computation results
-            /// </summary>
-            public Block8x8F WorkspaceBlock1;
-
-            /// <summary>
-            /// Temporal block 2 to store intermediate and/or final computation results
-            /// </summary>
-            public Block8x8F WorkspaceBlock2;
-
-            /// <summary>
-            /// The quantization table as <see cref="Block8x8F"/>
-            /// </summary>
-            public Block8x8F QuantiazationTable;
-
-            /// <summary>
-            /// The jpeg unzig data
-            /// </summary>
-            public UnzigData Unzig;
+            int qtIndex = component.QuantizationTableIndex;
+            this.DequantiazationTable = ZigZag.CreateDequantizationTable(ref decoder.QuantizationTables[qtIndex]);
+            this.subSamplingDivisors = component.SubSamplingDivisors;
 
-            /// <summary>
-            /// Creates and initializes a new <see cref="ComputationData"/> instance
-            /// </summary>
-            /// <returns>The <see cref="ComputationData"/></returns>
-            public static ComputationData Create()
-            {
-                var data = default(ComputationData);
-                data.Unzig = UnzigData.Create();
-                return data;
-            }
+            this.SourceBlock = default(Block8x8F);
+            this.WorkspaceBlock1 = default(Block8x8F);
+            this.WorkspaceBlock2 = default(Block8x8F);
         }
 
         /// <summary>
-        /// Contains pointers to the memory regions of <see cref="ComputationData"/> so they can be easily passed around to pointer based utility methods of <see cref="Block8x8F"/>
+        /// Processes 'sourceBlock' producing Jpeg color channel values from spectral values:
+        /// - Dequantize
+        /// - Applying IDCT
+        /// - Level shift by +128, clip to [0, 255]
+        /// - Copy the resultin color values into 'destArea' scaling up the block by amount defined in <see cref="subSamplingDivisors"/>
         /// </summary>
-        public struct DataPointers
+        public void ProcessBlockColorsInto(
+            ref Block8x8 sourceBlock,
+            BufferArea<float> destArea)
         {
-            /// <summary>
-            /// Pointer to <see cref="ComputationData.SourceBlock"/>
-            /// </summary>
-            public Block8x8F* SourceBlock;
-
-            /// <summary>
-            /// Pointer to <see cref="ComputationData.WorkspaceBlock1"/>
-            /// </summary>
-            public Block8x8F* WorkspaceBlock1;
+            ref Block8x8F b = ref this.SourceBlock;
+            b.LoadFrom(ref sourceBlock);
 
-            /// <summary>
-            /// Pointer to <see cref="ComputationData.WorkspaceBlock2"/>
-            /// </summary>
-            public Block8x8F* WorkspaceBlock2;
+            // Dequantize:
+            b.MultiplyInplace(ref this.DequantiazationTable);
 
-            /// <summary>
-            /// Pointer to <see cref="ComputationData.QuantiazationTable"/>
-            /// </summary>
-            public Block8x8F* QuantiazationTable;
+            FastFloatingPointDCT.TransformIDCT(ref b, ref this.WorkspaceBlock1, ref this.WorkspaceBlock2);
 
-            /// <summary>
-            /// Pointer to <see cref="ComputationData.Unzig"/> as int*
-            /// </summary>
-            public int* Unzig;
+            // To conform better to libjpeg we actually NEED TO loose precision here.
+            // This is because they store blocks as Int16 between all the operations.
+            // To be "more accurate", we need to emulate this by rounding!
+            this.WorkspaceBlock1.NormalizeColorsAndRoundInplace();
 
-            /// <summary>
-            /// Initializes a new instance of the <see cref="DataPointers" /> struct.
-            /// </summary>
-            /// <param name="dataPtr">Pointer to <see cref="ComputationData"/></param>
-            internal DataPointers(ComputationData* dataPtr)
-            {
-                this.SourceBlock = &dataPtr->SourceBlock;
-                this.WorkspaceBlock1 = &dataPtr->WorkspaceBlock1;
-                this.WorkspaceBlock2 = &dataPtr->WorkspaceBlock2;
-                this.QuantiazationTable = &dataPtr->QuantiazationTable;
-                this.Unzig = dataPtr->Unzig.Data;
-            }
+            this.WorkspaceBlock1.CopyTo(destArea, this.subSamplingDivisors.Width, this.subSamplingDivisors.Height);
         }
     }
 }

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

@@ -1,6 +1,7 @@
 using System;
 using System.Numerics;
 using System.Runtime.CompilerServices;
+using SixLabors.ImageSharp.Common.Tuples;
 
 namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
 {
@@ -121,9 +122,25 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
                     tmp.MultiplyInplace(1.772F);
                     b.AddInplace(ref tmp);
 
-                    r.RoundAndDownscale();
-                    g.RoundAndDownscale();
-                    b.RoundAndDownscale();
+                    if (Vector<float>.Count == 4)
+                    {
+                        // TODO: Find a way to properly run & test this path on AVX2 PC-s! (Have I already mentioned that Vector<T> is terrible?)
+                        r.RoundAndDownscalePreAvx2();
+                        g.RoundAndDownscalePreAvx2();
+                        b.RoundAndDownscalePreAvx2();
+                    }
+                    else if (SimdUtils.IsAvx2CompatibleArchitecture)
+                    {
+                        r.RoundAndDownscaleAvx2();
+                        g.RoundAndDownscaleAvx2();
+                        b.RoundAndDownscaleAvx2();
+                    }
+                    else
+                    {
+                        // TODO: Run fallback scalar code here
+                        // However, no issues expected before someone implements this: https://github.com/dotnet/coreclr/issues/12007
+                        throw new NotImplementedException("Your CPU architecture is too modern!");
+                    }
 
                     // 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);
@@ -131,53 +148,6 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
                 }
             }
 
-            /// <summary>
-            /// Its faster to process multiple Vector4-s together
-            /// </summary>
-            private struct Vector4Pair
-            {
-                public Vector4 A;
-
-                public Vector4 B;
-
-                private static readonly Vector4 Scale = new Vector4(1 / 255f);
-
-                private static readonly Vector4 Half = new Vector4(0.5f);
-
-                [MethodImpl(MethodImplOptions.AggressiveInlining)]
-                public void RoundAndDownscale()
-                {
-                    // Emulate rounding:
-                    this.A += Half;
-                    this.B += Half;
-
-                    // Downscale by 1/255
-                    this.A *= Scale;
-                    this.B *= Scale;
-                }
-
-                [MethodImpl(MethodImplOptions.AggressiveInlining)]
-                public void MultiplyInplace(float value)
-                {
-                    this.A *= value;
-                    this.B *= value;
-                }
-
-                [MethodImpl(MethodImplOptions.AggressiveInlining)]
-                public void AddInplace(Vector4 value)
-                {
-                    this.A += value;
-                    this.B += value;
-                }
-
-                [MethodImpl(MethodImplOptions.AggressiveInlining)]
-                public void AddInplace(ref Vector4Pair other)
-                {
-                    this.A += other.A;
-                    this.B += other.B;
-                }
-            }
-
             private struct Vector4Octet
             {
 #pragma warning disable SA1132 // Do not combine fields

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

@@ -66,10 +66,9 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
         /// <summary>
         /// Invoke <see cref="JpegBlockPostProcessor"/> for <see cref="BlockRowsPerStep"/> block rows, copy the result into <see cref="ColorBuffer"/>.
         /// </summary>
-        public unsafe void CopyBlocksToColorBuffer()
+        public void CopyBlocksToColorBuffer()
         {
-            var blockPp = default(JpegBlockPostProcessor);
-            JpegBlockPostProcessor.Init(&blockPp);
+            var blockPp = new JpegBlockPostProcessor(this.ImagePostProcessor.RawJpeg, this.Component);
 
             for (int y = 0; y < this.BlockRowsPerStep; y++)
             {
@@ -95,7 +94,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
                         this.blockAreaSize.Width,
                         this.blockAreaSize.Height);
 
-                    blockPp.ProcessBlockColorsInto(this.ImagePostProcessor.RawJpeg, this.Component, ref block, destArea);
+                    blockPp.ProcessBlockColorsInto(ref block, destArea);
                 }
             }
 

src/ImageSharp/Formats/Jpeg/Common/FastFloatingPointDCT.cs

@@ -50,7 +50,10 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
         /// <param name="temp">Temporary block provided by the caller</param>
         public static void TransformIDCT(ref Block8x8F src, ref Block8x8F dest, ref Block8x8F temp)
         {
+            // TODO: Transpose is a bottleneck now. We need full AVX support to optimize it:
+            // https://github.com/dotnet/corefx/issues/22940
             src.TransposeInto(ref temp);
+
             IDCT8x4_LeftPart(ref temp, ref dest);
             IDCT8x4_RightPart(ref temp, ref dest);
 
@@ -60,7 +63,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
             IDCT8x4_RightPart(ref temp, ref dest);
 
             // TODO: What if we leave the blocks in a scaled-by-x8 state until final color packing?
-            dest.MultiplyAllInplace(C_0_125);
+            dest.MultiplyInplace(C_0_125);
         }
 
         /// <summary>
@@ -334,7 +337,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
             FDCT8x4_LeftPart(ref temp, ref dest);
             FDCT8x4_RightPart(ref temp, ref dest);
 
-            dest.MultiplyAllInplace(C_0_125);
+            dest.MultiplyInplace(C_0_125);
         }
     }
 }

src/ImageSharp/Formats/Jpeg/Common/UnzigData.cs → src/ImageSharp/Formats/Jpeg/Common/ZigZag.cs

@@ -6,13 +6,12 @@ using System.Runtime.InteropServices;
 namespace SixLabors.ImageSharp.Formats.Jpeg.Common
 {
     /// <summary>
-    /// TODO: This should be simply just a <see cref="Block8x8"/>!
     /// Holds the Jpeg UnZig array in a value/stack type.
     /// Unzig maps from the zigzag ordering to the natural ordering. For example,
     /// unzig[3] is the column and row of the fourth element in zigzag order. The
     /// value is 16, which means first column (16%8 == 0) and third row (16/8 == 2).
     /// </summary>
-    internal unsafe struct UnzigData
+    internal unsafe struct ZigZag
     {
         /// <summary>
         /// Copy of <see cref="Unzig"/> in a value type
@@ -33,15 +32,30 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
             };
 
         /// <summary>
-        /// Creates and fills an instance of <see cref="UnzigData"/> with Jpeg unzig indices
+        /// Creates and fills an instance of <see cref="ZigZag"/> with Jpeg unzig indices
         /// </summary>
         /// <returns>The new instance</returns>
-        public static UnzigData Create()
+        public static ZigZag CreateUnzigTable()
         {
-            UnzigData result = default(UnzigData);
+            ZigZag result = default(ZigZag);
             int* unzigPtr = result.Data;
             Marshal.Copy(Unzig, 0, (IntPtr)unzigPtr, 64);
             return result;
         }
+
+        /// <summary>
+        /// Apply Zigging to the given quantization table, so it will be sufficient to multiply blocks for dequantizing them.
+        /// </summary>
+        public static Block8x8F CreateDequantizationTable(ref Block8x8F qt)
+        {
+            Block8x8F result = default(Block8x8F);
+
+            for (int i = 0; i < 64; i++)
+            {
+                result[Unzig[i]] = qt[i];
+            }
+
+            return result;
+        }
     }
 }

src/ImageSharp/Formats/Jpeg/GolangPort/Components/Decoder/OrigJpegScanDecoder.ComputationData.cs

@@ -26,7 +26,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.GolangPort.Components.Decoder
             /// <summary>
             /// The jpeg unzig data
             /// </summary>
-            public UnzigData Unzig;
+            public ZigZag Unzig;
 
             /// <summary>
             /// The buffer storing the <see cref="OrigComponentScan"/>-s for each component
@@ -45,7 +45,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.GolangPort.Components.Decoder
             public static ComputationData Create()
             {
                 ComputationData data = default(ComputationData);
-                data.Unzig = UnzigData.Create();
+                data.Unzig = ZigZag.CreateUnzigTable();
                 return data;
             }
         }

src/ImageSharp/Formats/Jpeg/GolangPort/JpegEncoderCore.cs

@@ -455,7 +455,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.GolangPort
             Block8x8F onStackLuminanceQuantTable = this.luminanceQuantTable;
             Block8x8F onStackChrominanceQuantTable = this.chrominanceQuantTable;
 
-            UnzigData unzig = UnzigData.Create();
+            ZigZag unzig = ZigZag.CreateUnzigTable();
 
             // ReSharper disable once InconsistentNaming
             int prevDCY = 0, prevDCCb = 0, prevDCCr = 0;
@@ -564,7 +564,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.GolangPort
         {
             FastFloatingPointDCT.TransformFDCT(ref *src, ref *tempDest1, ref *tempDest2);
 
-            Block8x8F.UnzigDivRound(tempDest1, tempDest2, quant, unzigPtr);
+            Block8x8F.Quantize(tempDest1, tempDest2, quant, unzigPtr);
             float* unziggedDestPtr = (float*)tempDest2;
 
             int dc = (int)unziggedDestPtr[0];
@@ -912,7 +912,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.GolangPort
             Block8x8F onStackLuminanceQuantTable = this.luminanceQuantTable;
             Block8x8F onStackChrominanceQuantTable = this.chrominanceQuantTable;
 
-            UnzigData unzig = UnzigData.Create();
+            ZigZag unzig = ZigZag.CreateUnzigTable();
 
             // ReSharper disable once InconsistentNaming
             int prevDCY = 0, prevDCCb = 0, prevDCCr = 0;

tests/ImageSharp.Tests/Common/SimdUtilsTests.cs

@@ -8,6 +8,8 @@ namespace SixLabors.ImageSharp.Tests.Common
     using System.Linq;
     using System.Runtime.CompilerServices;
 
+    using SixLabors.ImageSharp.Common.Tuples;
+
     using Xunit.Abstractions;
     using Xunit.Sdk;
 
@@ -243,15 +245,15 @@ namespace SixLabors.ImageSharp.Tests.Common
             
             x = (x * scale) + magick;
 
-            SimdUtils.Octet.OfUInt32 ii = default(SimdUtils.Octet.OfUInt32);
+            Tuple8.OfUInt32 ii = default(Tuple8.OfUInt32);
 
-            ref Vector<float> iiRef = ref Unsafe.As<SimdUtils.Octet.OfUInt32, Vector<float>>(ref ii);
+            ref Vector<float> iiRef = ref Unsafe.As<Tuple8.OfUInt32, Vector<float>>(ref ii);
 
             iiRef = x;
 
-            //SimdUtils.Octet.OfUInt32 ii = Unsafe.As<Vector<float>, SimdUtils.Octet.OfUInt32>(ref x);
+            //Tuple8.OfUInt32 ii = Unsafe.As<Vector<float>, Tuple8.OfUInt32>(ref x);
             
-            ref SimdUtils.Octet.OfByte d = ref dest.NonPortableCast<byte, SimdUtils.Octet.OfByte>()[0];
+            ref Tuple8.OfByte d = ref dest.NonPortableCast<byte, Tuple8.OfByte>()[0];
             d.LoadFrom(ref ii);
 
             this.Output.WriteLine(ii.ToString());

tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs

@@ -32,32 +32,18 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
         {
         }
 
-        [Fact]
-        public void Indexer()
+        private bool SkipOnNonAvx2Runner()
         {
-            float sum = 0;
-            this.Measure(
-                Times,
-                () =>
-                    {
-                        var block = new Block8x8F();
-
-                        for (int i = 0; i < Block8x8F.Size; i++)
-                        {
-                            block[i] = i;
-                        }
-
-                        sum = 0;
-                        for (int i = 0; i < Block8x8F.Size; i++)
-                        {
-                            sum += block[i];
-                        }
-                    });
-            Assert.Equal(sum, 64f * 63f * 0.5f);
+            if (!SimdUtils.IsAvx2CompatibleArchitecture)
+            {
+                this.Output.WriteLine("AVX2 not supported, skipping!");
+                return true;
+            }
+            return false;
         }
 
         [Fact]
-        public unsafe void Indexer_GetScalarAt_SetScalarAt()
+        public void Indexer()
         {
             float sum = 0;
             this.Measure(
@@ -68,18 +54,18 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
                         for (int i = 0; i < Block8x8F.Size; i++)
                         {
-                            Block8x8F.SetScalarAt(&block, i, i);
+                            block[i] = i;
                         }
 
                         sum = 0;
                         for (int i = 0; i < Block8x8F.Size; i++)
                         {
-                            sum += Block8x8F.GetScalarAt(&block, i);
+                            sum += block[i];
                         }
                     });
             Assert.Equal(sum, 64f * 63f * 0.5f);
         }
-
+        
         [Fact]
         public void Indexer_ReferenceBenchmarkWithArray()
         {
@@ -222,33 +208,6 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             this.Output.WriteLine($"TranposeInto_PinningImpl_Benchmark finished in {sw.ElapsedMilliseconds} ms");
         }
         
-        [Fact]
-        public unsafe void CopyColorsTo()
-        {
-            float[] data = Create8x8FloatData();
-            var block = new Block8x8F();
-            block.LoadFrom(data);
-            block.MultiplyAllInplace(5);
-
-            int stride = 256;
-            int height = 42;
-            int offset = height * 10 + 20;
-
-            byte[] colorsExpected = new byte[stride * height];
-            byte[] colorsActual = new byte[stride * height];
-
-            var temp = new Block8x8F();
-
-            ReferenceImplementations.CopyColorsTo(ref block, new Span<byte>(colorsExpected, offset), stride);
-
-            block.CopyColorsTo(new Span<byte>(colorsActual, offset), stride, &temp);
-
-            // Output.WriteLine("******* EXPECTED: *********");
-            // PrintLinearData(colorsExpected);
-            // Output.WriteLine("******** ACTUAL: **********");
-            Assert.Equal(colorsExpected, colorsActual);
-        }
-
         private static float[] Create8x8ColorCropTestData()
         {
             float[] result = new float[64];
@@ -263,30 +222,18 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             return result;
         }
 
-        [Theory]
-        [InlineData(false)]
-        [InlineData(true)]
-        public void NormalizeColors(bool inplace)
+        [Fact]
+        public void NormalizeColors()
         {
             var block = default(Block8x8F);
             float[] input = Create8x8ColorCropTestData();
             block.LoadFrom(input);
             this.Output.WriteLine("Input:");
             this.PrintLinearData(input);
-            
-            var dest = default(Block8x8F);
 
-            if (inplace)
-            {
-                dest = block;
-                dest.NormalizeColorsInplace();
-            }
-            else
-            {
-                block.NormalizeColorsInto(ref dest);
-            }
+            Block8x8F dest = block;
+            dest.NormalizeColorsInplace();
             
-
             float[] array = new float[64];
             dest.CopyTo(array);
             this.Output.WriteLine("Result:");
@@ -297,11 +244,35 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             }
         }
 
+        [Theory]
+        [InlineData(1)]
+        [InlineData(2)]
+        public void NormalizeColorsAndRoundAvx2(int seed)
+        {
+            if (this.SkipOnNonAvx2Runner())
+            {
+                return;
+            }
+
+            Block8x8F source = CreateRandomFloatBlock(-200, 200, seed);
+
+            Block8x8F expected = source;
+            expected.NormalizeColorsInplace();
+            expected.RoundInplace();
+
+            Block8x8F actual = source;
+            actual.NormalizeColorsAndRoundInplaceAvx2();
+
+            this.Output.WriteLine(expected.ToString());
+            this.Output.WriteLine(actual.ToString());
+            this.CompareBlocks(expected, actual, 0);
+        }
+
 
         [Theory]
         [InlineData(1)]
         [InlineData(2)]
-        public unsafe void UnzigDivRound(int seed)
+        public unsafe void Quantize(int seed)
         {
             var block = new Block8x8F();
             block.LoadFrom(Create8x8RoundedRandomFloatData(-2000, 2000, seed));
@@ -309,14 +280,14 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             var qt = new Block8x8F();
             qt.LoadFrom(Create8x8RoundedRandomFloatData(-2000, 2000, seed));
 
-            var unzig = UnzigData.Create();
+            var unzig = ZigZag.CreateUnzigTable();
 
             int* expectedResults = stackalloc int[Block8x8F.Size];
-            ReferenceImplementations.UnZigDivRoundRational(&block, expectedResults, &qt, unzig.Data);
+            ReferenceImplementations.QuantizeRational(&block, expectedResults, &qt, unzig.Data);
 
             var actualResults = default(Block8x8F);
 
-            Block8x8F.UnzigDivRound(&block, &actualResults, &qt, unzig.Data);
+            Block8x8F.Quantize(&block, &actualResults, &qt, unzig.Data);
 
             for (int i = 0; i < Block8x8F.Size; i++)
             {
@@ -352,7 +323,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
         [InlineData(1)]
         [InlineData(2)]
         [InlineData(3)]
-        public void RoundInplace(int seed)
+        public void RoundInplaceSlow(int seed)
         {
             Block8x8F s = CreateRandomFloatBlock(-500, 500, seed);
 
@@ -370,5 +341,77 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 Assert.Equal(expected, actual);
             }
         }
+
+        [Fact]
+        public void MultiplyInplace_ByOtherBlock()
+        {
+            Block8x8F original = CreateRandomFloatBlock(-500, 500, 42);
+            Block8x8F m = CreateRandomFloatBlock(-500, 500, 42);
+
+            Block8x8F actual = original;
+            
+            actual.MultiplyInplace(ref m);
+
+            for (int i = 0; i < Block8x8F.Size; i++)
+            {
+                Assert.Equal(original[i]*m[i], actual[i]);
+            }
+        }
+
+        [Theory]
+        [InlineData(1)]
+        [InlineData(2)]
+        [InlineData(3)]
+        public unsafe void DequantizeBlock(int seed)
+        {
+            Block8x8F original = CreateRandomFloatBlock(-500, 500, seed);
+            Block8x8F qt = CreateRandomFloatBlock(0, 10, seed + 42);
+
+            var unzig = ZigZag.CreateUnzigTable();
+
+            Block8x8F expected = original;
+            Block8x8F actual = original;
+
+            ReferenceImplementations.DequantizeBlock(&expected, &qt, unzig.Data);
+            Block8x8F.DequantizeBlock(&actual, &qt, unzig.Data);
+
+            this.CompareBlocks(expected, actual, 0);
+        }
+
+        [Theory]
+        [InlineData(1)]
+        [InlineData(2)]
+        [InlineData(3)]
+        public unsafe void ZigZag_CreateDequantizationTable_MultiplicationShouldQuantize(int seed)
+        {
+            Block8x8F original = CreateRandomFloatBlock(-500, 500, seed);
+            Block8x8F qt = CreateRandomFloatBlock(0, 10, seed + 42);
+
+            var unzig = ZigZag.CreateUnzigTable();
+            Block8x8F zigQt = ZigZag.CreateDequantizationTable(ref qt);
+
+            Block8x8F expected = original;
+            Block8x8F actual = original;
+
+            ReferenceImplementations.DequantizeBlock(&expected, &qt, unzig.Data);
+
+            actual.MultiplyInplace(ref zigQt);
+
+            this.CompareBlocks(expected, actual, 0);
+        }
+
+        [Fact]
+        public void MultiplyInplace_ByScalar()
+        {
+            Block8x8F original = CreateRandomFloatBlock(-500, 500);
+
+            Block8x8F actual = original;
+            actual.MultiplyInplace(42f);
+
+            for (int i = 0; i < 64; i++)
+            {
+                Assert.Equal(original[i]*42f, actual[i]);
+            }
+        }
     }
 }

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

@@ -14,7 +14,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 {
     public class JpegColorConverterTests
     {
-        private const float Precision = 1/255f;
+        private const float Precision = 0.1f / 255;
 
         public static readonly TheoryData<int, int, int> CommonConversionData =
             new TheoryData<int, int, int>

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

@@ -9,6 +9,7 @@ using System;
 
 namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 {
+    using System.Collections.Generic;
     using System.IO;
     using System.Linq;
 
@@ -24,6 +25,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
     using Xunit;
     using Xunit.Abstractions;
 
+    // TODO: Scatter test cases into multiple test classes
     public class JpegDecoderTests
     {
         public static string[] BaselineTestJpegs =
@@ -50,16 +52,43 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 TestImages.Jpeg.Issues.MissingFF00ProgressiveGirl159,
             };
 
+        private static readonly Dictionary<string, float> CustomToleranceValues = new Dictionary<string, float>
+        {
+            // Baseline:
+            [TestImages.Jpeg.Baseline.Calliphora] = 0.00002f / 100,
+            [TestImages.Jpeg.Baseline.Bad.ExifUndefType] = 0.011f / 100,
+            [TestImages.Jpeg.Baseline.Bad.BadEOF] = 0.38f / 100,
+            [TestImages.Jpeg.Baseline.Testorig420] = 0.38f / 100,
+
+            // Progressive:
+            [TestImages.Jpeg.Issues.MissingFF00ProgressiveGirl159] = 0.34f / 100,
+            [TestImages.Jpeg.Issues.BadCoeffsProgressive178] = 0.38f / 100,
+            [TestImages.Jpeg.Progressive.Bad.BadEOF] = 0.3f / 100,
+            [TestImages.Jpeg.Progressive.Festzug] = 0.02f / 100,
+            [TestImages.Jpeg.Progressive.Fb] = 0.16f / 100,
+            [TestImages.Jpeg.Progressive.Progress] = 0.31f / 100,
+        };
+
         public const PixelTypes CommonNonDefaultPixelTypes = PixelTypes.Rgba32 | PixelTypes.Argb32 | PixelTypes.RgbaVector;
 
-        // TODO: We should make this comparer less tolerant ...
-        private static readonly ImageComparer VeryTolerantJpegComparer =
-            ImageComparer.Tolerant(0.005f, perPixelManhattanThreshold: 4);
+        private const float BaselineTolerance_Orig = 0.001f / 100;
+        private const float BaselineTolerance_PdfJs = 0.005f;
 
-        // BUG: PDF.js output is wrong on spectral level!
-        private static readonly ImageComparer PdfJsProgressiveComparer =
-            ImageComparer.Tolerant(0.015f, perPixelManhattanThreshold: 4);
+        private const float ProgressiveTolerance_Orig = 0.2f / 100;
+        private const float ProgressiveTolerance_PdfJs = 1.5f / 100; // PDF.js Progressive output is wrong on spectral level!
 
+        private ImageComparer GetImageComparerForOrigDecoder<TPixel>(TestImageProvider<TPixel> provider)
+            where TPixel : struct, IPixel<TPixel>
+        {
+            string file = provider.SourceFileOrDescription;
+
+            if (!CustomToleranceValues.TryGetValue(file, out float tolerance))
+            {
+                tolerance = file.ToLower().Contains("baseline") ? BaselineTolerance_Orig : ProgressiveTolerance_Orig;
+            }
+
+            return ImageComparer.Tolerant(tolerance);
+        }
 
         public JpegDecoderTests(ITestOutputHelper output)
         {
@@ -71,7 +100,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
         private static IImageDecoder OrigJpegDecoder => new OrigJpegDecoder();
 
         private static IImageDecoder PdfJsJpegDecoder => new PdfJsJpegDecoder();
-        
+
         [Fact]
         public void ParseStream_BasicPropertiesAreCorrect1_PdfJs()
         {
@@ -84,50 +113,56 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 VerifyJpeg.VerifyComponentSizes3(decoder.Frame.Components, 43, 61, 22, 31, 22, 31);
             }
         }
-        
+
         public const string DecodeBaselineJpegOutputName = "DecodeBaselineJpeg";
 
+
         [Theory]
-        [WithFileCollection(nameof(BaselineTestJpegs), PixelTypes.Rgba32)]
-        public void DecodeBaselineJpeg_PdfJs<TPixel>(TestImageProvider<TPixel> provider)
+        [WithFile(TestImages.Jpeg.Baseline.Calliphora, CommonNonDefaultPixelTypes, false)]
+        [WithFile(TestImages.Jpeg.Baseline.Calliphora, CommonNonDefaultPixelTypes, true)]
+        public void JpegDecoder_IsNotBoundToSinglePixelType<TPixel>(TestImageProvider<TPixel> provider, bool useOldDecoder)
             where TPixel : struct, IPixel<TPixel>
         {
-            using (Image<TPixel> image = provider.GetImage(PdfJsJpegDecoder))
+            IImageDecoder decoder = useOldDecoder ? OrigJpegDecoder : PdfJsJpegDecoder;
+            using (Image<TPixel> image = provider.GetImage(decoder))
             {
                 image.DebugSave(provider);
 
                 provider.Utility.TestName = DecodeBaselineJpegOutputName;
-                image.CompareToReferenceOutput(provider, VeryTolerantJpegComparer, appendPixelTypeToFileName: false);
+                image.CompareToReferenceOutput(provider, ImageComparer.Tolerant(BaselineTolerance_PdfJs), appendPixelTypeToFileName: false);
             }
         }
 
         [Theory]
-        [WithFile(TestImages.Jpeg.Baseline.Calliphora, CommonNonDefaultPixelTypes, false)]
-        [WithFile(TestImages.Jpeg.Baseline.Calliphora, CommonNonDefaultPixelTypes, true)]
-        public void JpegDecoder_IsNotBoundToSinglePixelType<TPixel>(TestImageProvider<TPixel> provider, bool useOldDecoder)
+        [WithFileCollection(nameof(BaselineTestJpegs), PixelTypes.Rgba32)]
+        public void DecodeBaselineJpeg_Orig<TPixel>(TestImageProvider<TPixel> provider)
             where TPixel : struct, IPixel<TPixel>
         {
-            IImageDecoder decoder = useOldDecoder ? OrigJpegDecoder : PdfJsJpegDecoder;
-            using (Image<TPixel> image = provider.GetImage(decoder))
+            using (Image<TPixel> image = provider.GetImage(OrigJpegDecoder))
             {
                 image.DebugSave(provider);
-
                 provider.Utility.TestName = DecodeBaselineJpegOutputName;
-                image.CompareToReferenceOutput(provider, VeryTolerantJpegComparer, appendPixelTypeToFileName: false);
+                image.CompareToReferenceOutput(
+                    provider,
+                    this.GetImageComparerForOrigDecoder(provider),
+                    appendPixelTypeToFileName: false);
             }
         }
-        
+
         [Theory]
         [WithFileCollection(nameof(BaselineTestJpegs), PixelTypes.Rgba32)]
-        public void DecodeBaselineJpeg_Orig<TPixel>(TestImageProvider<TPixel> provider)
+        public void DecodeBaselineJpeg_PdfJs<TPixel>(TestImageProvider<TPixel> provider)
             where TPixel : struct, IPixel<TPixel>
         {
-            using (Image<TPixel> image = provider.GetImage(OrigJpegDecoder))
+            using (Image<TPixel> image = provider.GetImage(PdfJsJpegDecoder))
             {
                 image.DebugSave(provider);
 
                 provider.Utility.TestName = DecodeBaselineJpegOutputName;
-                image.CompareToReferenceOutput(provider, VeryTolerantJpegComparer, appendPixelTypeToFileName: false);
+                image.CompareToReferenceOutput(
+                    provider,
+                    ImageComparer.Tolerant(BaselineTolerance_PdfJs),
+                    appendPixelTypeToFileName: false);
             }
         }
 
@@ -144,37 +179,43 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
         [Theory]
         [WithFileCollection(nameof(ProgressiveTestJpegs), PixelTypes.Rgba32)]
-        public void DecodeProgressiveJpeg_PdfJs<TPixel>(TestImageProvider<TPixel> provider)
+        public void DecodeProgressiveJpeg_Orig<TPixel>(TestImageProvider<TPixel> provider)
             where TPixel : struct, IPixel<TPixel>
         {
-            using (Image<TPixel> image = provider.GetImage(PdfJsJpegDecoder))
+            using (Image<TPixel> image = provider.GetImage(OrigJpegDecoder))
             {
                 image.DebugSave(provider);
 
                 provider.Utility.TestName = DecodeProgressiveJpegOutputName;
-                image.CompareToReferenceOutput(provider, PdfJsProgressiveComparer, appendPixelTypeToFileName: false);
+                image.CompareToReferenceOutput(
+                    provider,
+                    this.GetImageComparerForOrigDecoder(provider),
+                    appendPixelTypeToFileName: false);
             }
         }
 
         [Theory]
         [WithFileCollection(nameof(ProgressiveTestJpegs), PixelTypes.Rgba32)]
-        public void DecodeProgressiveJpeg_Orig<TPixel>(TestImageProvider<TPixel> provider)
+        public void DecodeProgressiveJpeg_PdfJs<TPixel>(TestImageProvider<TPixel> provider)
             where TPixel : struct, IPixel<TPixel>
         {
-            using (Image<TPixel> image = provider.GetImage(OrigJpegDecoder))
+            using (Image<TPixel> image = provider.GetImage(PdfJsJpegDecoder))
             {
                 image.DebugSave(provider);
 
                 provider.Utility.TestName = DecodeProgressiveJpegOutputName;
-                image.CompareToReferenceOutput(provider, VeryTolerantJpegComparer, appendPixelTypeToFileName: false);
+                image.CompareToReferenceOutput(
+                    provider,
+                    ImageComparer.Tolerant(ProgressiveTolerance_PdfJs),
+                    appendPixelTypeToFileName: false);
             }
         }
-        
-        private float GetDifferenceInPercents<TPixel>(Image<TPixel> image, TestImageProvider<TPixel> provider)
+
+        private string GetDifferenceInPercentageString<TPixel>(Image<TPixel> image, TestImageProvider<TPixel> provider)
             where TPixel : struct, IPixel<TPixel>
         {
             var reportingComparer = ImageComparer.Tolerant(0, 0);
-            
+
             ImageSimilarityReport report = image.GetReferenceOutputSimilarityReports(
                 provider,
                 reportingComparer,
@@ -183,10 +224,10 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
             if (report != null && report.TotalNormalizedDifference.HasValue)
             {
-                return report.TotalNormalizedDifference.Value * 100;
+                return report.DifferencePercentageString;
             }
 
-            return 0;
+            return "0%";
         }
 
         private void CompareJpegDecodersImpl<TPixel>(TestImageProvider<TPixel> provider, string testName)
@@ -202,14 +243,15 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
             using (Image<TPixel> image = provider.GetImage(OrigJpegDecoder))
             {
-                double d = this.GetDifferenceInPercents(image, provider);
-                this.Output.WriteLine($"Difference using ORIGINAL decoder: {d:0.0000}%");
+                string d = this.GetDifferenceInPercentageString(image, provider);
+
+                this.Output.WriteLine($"Difference using ORIGINAL decoder: {d}");
             }
 
             using (Image<TPixel> image = provider.GetImage(PdfJsJpegDecoder))
             {
-                double d = this.GetDifferenceInPercents(image, provider);
-                this.Output.WriteLine($"Difference using PDFJS decoder: {d:0.0000}%");
+                string d = this.GetDifferenceInPercentageString(image, provider);
+                this.Output.WriteLine($"Difference using PDFJS decoder: {d}");
             }
         }
 
@@ -228,7 +270,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
         {
             this.CompareJpegDecodersImpl(provider, DecodeProgressiveJpegOutputName);
         }
-        
+
         [Theory]
         [WithSolidFilledImages(16, 16, 255, 0, 0, PixelTypes.Rgba32, JpegSubsample.Ratio420, 75)]
         [WithSolidFilledImages(16, 16, 255, 0, 0, PixelTypes.Rgba32, JpegSubsample.Ratio420, 100)]
@@ -256,7 +298,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             var mirror = Image.Load<TPixel>(data, OrigJpegDecoder);
             mirror.DebugSave(provider, $"_{subsample}_Q{quality}");
         }
-        
+
         [Fact]
         public void Decoder_Reads_Correct_Resolution_From_Jfif()
         {
@@ -314,7 +356,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
         // into "\tests\Images\ActualOutput\JpegDecoderTests\"
         //[Theory]
         //[WithFile(TestImages.Jpeg.Progressive.Progress, PixelTypes.Rgba32, "PdfJsOriginal_progress.png")]
-        public void ValidateProgressivePdfJsOutput<TPixel>(TestImageProvider<TPixel> provider, 
+        public void ValidateProgressivePdfJsOutput<TPixel>(TestImageProvider<TPixel> provider,
             string pdfJsOriginalResultImage)
             where TPixel : struct, IPixel<TPixel>
         {
@@ -335,7 +377,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             {
                 ImageSimilarityReport originalReport = comparer.CompareImagesOrFrames(expectedImage, pdfJsOriginalResult);
                 ImageSimilarityReport portReport = comparer.CompareImagesOrFrames(expectedImage, pdfJsPortResult);
-                
+
                 this.Output.WriteLine($"Difference for PDF.js ORIGINAL: {originalReport.DifferencePercentageString}");
                 this.Output.WriteLine($"Difference for PORT: {portReport.DifferencePercentageString}");
             }

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

@@ -34,8 +34,8 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             TestImages.Jpeg.Baseline.Jpeg444,
         };
 
-        // [Theory] // Benchmark, enable manually
-        // [MemberData(nameof(DecodeJpegData))]
+        //[Theory] // Benchmark, enable manually
+        //[MemberData(nameof(DecodeJpegData))]
         public void DecodeJpeg_Original(string fileName)
         {
             this.DecodeJpegBenchmarkImpl(fileName, new OrigJpegDecoder());

tests/ImageSharp.Tests/Formats/Jpg/Utils/ReferenceImplementations.cs

@@ -19,6 +19,21 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg.Utils
     /// </summary>
     internal static partial class ReferenceImplementations
     {
+        public static unsafe void DequantizeBlock(Block8x8F* blockPtr, Block8x8F* qtPtr, int* unzigPtr)
+        {
+            float* b = (float*)blockPtr;
+            float* qtp = (float*)qtPtr;
+            for (int qtIndex = 0; qtIndex < Block8x8F.Size; qtIndex++)
+            {
+                int i = unzigPtr[qtIndex];
+                float* unzigPos = b + i;
+
+                float val = *unzigPos;
+                val *= qtp[qtIndex];
+                *unzigPos = val;
+            }
+        }
+
         /// <summary>
         /// Transpose 8x8 block stored linearly in a <see cref="Span{T}"/> (inplace)
         /// </summary>
@@ -93,14 +108,14 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg.Utils
         }
 
         /// <summary>
-        /// Reference implementation to test <see cref="Block8x8F.UnzigDivRound"/>.
+        /// Reference implementation to test <see cref="Block8x8F.Quantize"/>.
         /// Rounding is done used an integer-based algorithm defined in <see cref="RationalRound(int,int)"/>.
         /// </summary>
         /// <param name="src">The input block</param>
         /// <param name="dest">The destination block of integers</param>
         /// <param name="qt">The quantization table</param>
-        /// <param name="unzigPtr">Pointer to <see cref="UnzigData.Data"/> </param>
-        public static unsafe void UnZigDivRoundRational(Block8x8F* src, int* dest, Block8x8F* qt, int* unzigPtr)
+        /// <param name="unzigPtr">Pointer to <see cref="ZigZag.Data"/> </param>
+        public static unsafe void QuantizeRational(Block8x8F* src, int* dest, Block8x8F* qt, int* unzigPtr)
         {
             float* s = (float*)src;
             float* q = (float*)qt;

tests/ImageSharp.Tests/Processing/Processors/Transforms/ResizeTests.cs

@@ -44,7 +44,7 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Transforms
             {
                 SizeF newSize = image.Size() * ratio;
                 image.Mutate(x => x.Resize((Size)newSize, sampler, false));
-                string details = $"{name}-{ratio}";
+                string details = $"{name}-{ratio.ToString(System.Globalization.CultureInfo.InvariantCulture)}";
 
                 image.DebugSave(provider, details);
                 image.CompareToReferenceOutput(provider, details);

tests/ImageSharp.Tests/TestUtilities/ApproximateFloatComparer.cs

@@ -20,7 +20,7 @@ namespace SixLabors.ImageSharp.Tests
         {
             float d = x - y;
 
-            return d > -this.Eps && d < this.Eps;
+            return d >= -this.Eps && d <= this.Eps;
         }
 
         public int GetHashCode(float obj)

tests/ImageSharp.Tests/TestUtilities/ImageComparison/ImageSimilarityReport.cs

@@ -26,9 +26,24 @@
         // TODO: This should not be a nullable value!
         public float? TotalNormalizedDifference { get; }
 
-        public string DifferencePercentageString => this.TotalNormalizedDifference.HasValue
-                                                  ? $"{this.TotalNormalizedDifference.Value * 100:0.0000}%"
-                                                  : "?";
+        public string DifferencePercentageString
+        {
+            get
+            {
+                if (!this.TotalNormalizedDifference.HasValue)
+                {
+                    return "?";
+                }
+                else if (this.TotalNormalizedDifference == 0)
+                {
+                    return "0%";
+                }
+                else
+                {
+                    return $"{this.TotalNormalizedDifference.Value * 100:0.0000}%";
+                }
+            }
+        }
 
         public PixelDifference[] Differences { get; }
 

tests/ImageSharp.Tests/TestUtilities/ImageProviders/FileProvider.cs

@@ -105,15 +105,20 @@ namespace SixLabors.ImageSharp.Tests
 
             private static readonly ConcurrentDictionary<Key, Image<TPixel>> cache = new ConcurrentDictionary<Key, Image<TPixel>>();
 
-            public FileProvider(string filePath)
+            // Needed for deserialization!
+            // ReSharper disable once UnusedMember.Local
+            public FileProvider()
             {
-                this.FilePath = filePath;
             }
 
-            public FileProvider()
+            public FileProvider(string filePath)
             {
+                this.FilePath = filePath;
             }
-
+            
+            /// <summary>
+            /// Gets the file path relative to the "~/tests/images" folder
+            /// </summary>
             public string FilePath { get; private set; }
 
             public override string SourceFileOrDescription => this.FilePath;

tests/ImageSharp.Tests/TestUtilities/ImagingTestCaseUtility.cs

@@ -89,6 +89,8 @@ namespace SixLabors.ImageSharp.Tests
             return $"{this.GetTestOutputDir()}/{this.TestName}{pixName}{fn}{details}{extension}";
         }
 
+        private static string Inv(FormattableString formattable) => System.FormattableString.Invariant(formattable);
+        
         /// <summary>
         /// Gets the recommended file name for the output of the test
         /// </summary>
@@ -111,13 +113,17 @@ namespace SixLabors.ImageSharp.Tests
                 TypeInfo info = type.GetTypeInfo();
                 if (info.IsPrimitive || info.IsEnum || type == typeof(decimal))
                 {
-                    detailsString = testOutputDetails.ToString();
+                    detailsString = Inv($"{testOutputDetails}");
                 }
                 else
                 {
                     IEnumerable<PropertyInfo> properties = testOutputDetails.GetType().GetRuntimeProperties();
 
-                    detailsString = String.Join("_", properties.ToDictionary(x => x.Name, x => x.GetValue(testOutputDetails)).Select(x => $"{x.Key}-{x.Value}"));
+                    detailsString = string.Join(
+                        "_",
+                        properties.ToDictionary(x => x.Name, x => x.GetValue(testOutputDetails))
+                            .Select(x => Inv($"{x.Key}-{x.Value}"))
+                        );
                 }
             }
             return this.GetTestOutputFileNameImpl(extension, detailsString, appendPixelTypeToFileName);