JpegEncoder - Optimize Some Low Hanging Fruit

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

@@ -18,6 +18,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
     /// <summary>
     /// Represents a Jpeg block with <see cref="float"/> coefficients.
     /// </summary>
+    [StructLayout(LayoutKind.Sequential)]
     internal partial struct Block8x8F : IEquatable<Block8x8F>
     {
         /// <summary>
@@ -51,9 +52,6 @@ internal partial struct Block8x8F : IEquatable<Block8x8F>
         public Vector4 V7R;
 #pragma warning restore SA1600 // ElementsMustBeDocumented
 
-        private static readonly Vector4 NegativeOne = new Vector4(-1);
-        private static readonly Vector4 Offset = new Vector4(.5F);
-
         /// <summary>
         /// Get/Set scalar elements at a given index
         /// </summary>
@@ -155,10 +153,7 @@ public static Block8x8F Load(Span<int> data)
         /// </summary>
         [MethodImpl(InliningOptions.ShortMethod)]
         public void Clear()
-        {
-            // The cheapest way to do this in C#:
-            this = default;
-        }
+            => this = default; // The cheapest way to do this in C#:
 
         /// <summary>
         /// Load raw 32bit floating point data from source.
@@ -180,9 +175,7 @@ public void LoadFrom(Span<float> source)
         /// <param name="source">Source</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static unsafe void LoadFrom(Block8x8F* blockPtr, Span<float> source)
-        {
-            blockPtr->LoadFrom(source);
-        }
+            => blockPtr->LoadFrom(source);
 
         /// <summary>
         /// Load raw 32bit floating point data from source
@@ -236,9 +229,7 @@ public static unsafe void ScaledCopyTo(Block8x8F* blockPtr, Span<byte> dest)
         /// <param name="dest">The destination.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
         public static unsafe void ScaledCopyTo(Block8x8F* blockPtr, Span<float> dest)
-        {
-            blockPtr->ScaledCopyTo(dest);
-        }
+            => blockPtr->ScaledCopyTo(dest);
 
         /// <summary>
         /// Copy raw 32bit floating point data to dest
@@ -439,7 +430,6 @@ public void AddInPlace(float value)
         /// <param name="blockPtr">The block pointer.</param>
         /// <param name="qtPtr">The qt pointer.</param>
         /// <param name="unzigPtr">Unzig pointer</param>
-        // [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static unsafe void DequantizeBlock(Block8x8F* blockPtr, Block8x8F* qtPtr, byte* unzigPtr)
         {
             float* b = (float*)blockPtr;
@@ -556,22 +546,60 @@ private static unsafe void Scale16X16To8X8Scalar(ref Block8x8F destination, Read
         [MethodImpl(InliningOptions.ShortMethod)]
         private static void DivideRoundAll(ref Block8x8F a, ref Block8x8F b)
         {
-            a.V0L = DivideRound(a.V0L, b.V0L);
-            a.V0R = DivideRound(a.V0R, b.V0R);
-            a.V1L = DivideRound(a.V1L, b.V1L);
-            a.V1R = DivideRound(a.V1R, b.V1R);
-            a.V2L = DivideRound(a.V2L, b.V2L);
-            a.V2R = DivideRound(a.V2R, b.V2R);
-            a.V3L = DivideRound(a.V3L, b.V3L);
-            a.V3R = DivideRound(a.V3R, b.V3R);
-            a.V4L = DivideRound(a.V4L, b.V4L);
-            a.V4R = DivideRound(a.V4R, b.V4R);
-            a.V5L = DivideRound(a.V5L, b.V5L);
-            a.V5R = DivideRound(a.V5R, b.V5R);
-            a.V6L = DivideRound(a.V6L, b.V6L);
-            a.V6R = DivideRound(a.V6R, b.V6R);
-            a.V7L = DivideRound(a.V7L, b.V7L);
-            a.V7R = DivideRound(a.V7R, b.V7R);
+#if SUPPORTS_RUNTIME_INTRINSICS
+            if (Avx.IsSupported)
+            {
+                var vnegOne = Vector256.Create(-1f);
+                var vadd = Vector256.Create(.5F);
+                var vone = Vector256.Create(1f);
+
+                ref Vector256<float> aBase = ref Unsafe.AsRef(Unsafe.As<Vector4, Vector256<float>>(ref a.V0L));
+                ref Vector256<float> bBase = ref Unsafe.AsRef(Unsafe.As<Vector4, Vector256<float>>(ref b.V0L));
+                ref Vector256<float> aEnd = ref Unsafe.Add(ref aBase, 8);
+
+                do
+                {
+                    Vector256<float> voff = Avx.Multiply(Avx.Min(Avx.Max(vnegOne, aBase), vone), vadd);
+                    Unsafe.Add(ref aBase, 0) = Avx.Add(Avx.Divide(aBase, bBase), voff);
+
+                    aBase = ref Unsafe.Add(ref aBase, 1);
+                    bBase = ref Unsafe.Add(ref bBase, 1);
+                }
+                while (Unsafe.IsAddressLessThan(ref aBase, ref aEnd));
+            }
+            else
+#endif
+            {
+                a.V0L = DivideRound(a.V0L, b.V0L);
+                a.V0R = DivideRound(a.V0R, b.V0R);
+                a.V1L = DivideRound(a.V1L, b.V1L);
+                a.V1R = DivideRound(a.V1R, b.V1R);
+                a.V2L = DivideRound(a.V2L, b.V2L);
+                a.V2R = DivideRound(a.V2R, b.V2R);
+                a.V3L = DivideRound(a.V3L, b.V3L);
+                a.V3R = DivideRound(a.V3R, b.V3R);
+                a.V4L = DivideRound(a.V4L, b.V4L);
+                a.V4R = DivideRound(a.V4R, b.V4R);
+                a.V5L = DivideRound(a.V5L, b.V5L);
+                a.V5R = DivideRound(a.V5R, b.V5R);
+                a.V6L = DivideRound(a.V6L, b.V6L);
+                a.V6R = DivideRound(a.V6R, b.V6R);
+                a.V7L = DivideRound(a.V7L, b.V7L);
+                a.V7R = DivideRound(a.V7R, b.V7R);
+            }
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static Vector4 DivideRound(Vector4 dividend, Vector4 divisor)
+        {
+            var neg = new Vector4(-1);
+            var add = new Vector4(.5F);
+
+            // sign(dividend) = max(min(dividend, 1), -1)
+            Vector4 sign = Numerics.Clamp(dividend, neg, Vector4.One);
+
+            // AlmostRound(dividend/divisor) = dividend/divisor + 0.5*sign(dividend)
+            return (dividend / divisor) + (sign * add);
         }
 
         public void RoundInto(ref Block8x8 dest)
@@ -673,8 +701,7 @@ public void LoadFromInt16ExtendedAvx2(ref Block8x8 source)
 
         /// <inheritdoc />
         public bool Equals(Block8x8F other)
-        {
-            return this.V0L == other.V0L
+            => this.V0L == other.V0L
             && this.V0R == other.V0R
             && this.V1L == other.V1L
             && this.V1R == other.V1R
@@ -690,7 +717,6 @@ public bool Equals(Block8x8F other)
             && this.V6R == other.V6R
             && this.V7L == other.V7L
             && this.V7R == other.V7R;
-        }
 
         /// <inheritdoc />
         public override string ToString()
@@ -718,16 +744,6 @@ private static Vector<float> NormalizeAndRound(Vector<float> row, Vector<float>
             return row.FastRound();
         }
 
-        [MethodImpl(InliningOptions.ShortMethod)]
-        private static Vector4 DivideRound(Vector4 dividend, Vector4 divisor)
-        {
-            // sign(dividend) = max(min(dividend, 1), -1)
-            Vector4 sign = Numerics.Clamp(dividend, NegativeOne, Vector4.One);
-
-            // AlmostRound(dividend/divisor) = dividend/divisor + 0.5*sign(dividend)
-            return (dividend / divisor) + (sign * Offset);
-        }
-
         [Conditional("DEBUG")]
         private static void GuardBlockIndex(int idx)
         {

src/ImageSharp/Formats/Jpeg/Components/Encoder/YCbCrForwardConverter{TPixel}.cs

@@ -59,9 +59,9 @@ public static YCbCrForwardConverter<TPixel> Create()
         /// <summary>
         /// Converts a 8x8 image area inside 'pixels' at position (x,y) placing the result members of the structure (<see cref="Y"/>, <see cref="Cb"/>, <see cref="Cr"/>)
         /// </summary>
-        public void Convert(ImageFrame<TPixel> frame, int x, int y, in RowOctet<TPixel> currentRows)
+        public void Convert(ImageFrame<TPixel> frame, int x, int y, ref RowOctet<TPixel> currentRows)
         {
-            this.pixelBlock.LoadAndStretchEdges(frame.PixelBuffer, x, y, currentRows);
+            this.pixelBlock.LoadAndStretchEdges(frame.PixelBuffer, x, y, ref currentRows);
 
             Span<Rgb24> rgbSpan = this.rgbBlock.AsSpanUnsafe();
             PixelOperations<TPixel>.Instance.ToRgb24(frame.GetConfiguration(), this.pixelBlock.AsSpanUnsafe(), rgbSpan);
@@ -76,7 +76,7 @@ public void Convert(ImageFrame<TPixel> frame, int x, int y, in RowOctet<TPixel>
             }
             else
             {
-                this.colorTables.Convert(rgbSpan,  ref yBlock, ref cbBlock, ref crBlock);
+                this.colorTables.Convert(rgbSpan, ref yBlock, ref cbBlock, ref crBlock);
             }
         }
     }

src/ImageSharp/Formats/Jpeg/Components/GenericBlock8x8.cs

@@ -57,7 +57,7 @@ public T this[int idx]
         /// Load a 8x8 region of an image into the block.
         /// The "outlying" area of the block will be stretched out with pixels on the right and bottom edge of the image.
         /// </summary>
-        public void LoadAndStretchEdges(Buffer2D<T> source, int sourceX, int sourceY, in RowOctet<T> currentRows)
+        public void LoadAndStretchEdges(Buffer2D<T> source, int sourceX, int sourceY, ref RowOctet<T> currentRows)
         {
             int width = Math.Min(8, source.Width - sourceX);
             int height = Math.Min(8, source.Height - sourceY);

src/ImageSharp/Formats/Jpeg/Components/RowOctet.cs

@@ -1,4 +1,4 @@
-// Copyright (c) Six Labors.
+// Copyright (c) Six Labors.
 // Licensed under the Apache License, Version 2.0.
 
 using System;
@@ -12,39 +12,24 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
     /// Cache 8 pixel rows on the stack, which may originate from different buffers of a <see cref="MemoryGroup{T}"/>.
     /// </summary>
     [StructLayout(LayoutKind.Sequential)]
-    internal readonly ref struct RowOctet<T>
+    internal ref struct RowOctet<T>
         where T : struct
     {
-        private readonly Span<T> row0;
-        private readonly Span<T> row1;
-        private readonly Span<T> row2;
-        private readonly Span<T> row3;
-        private readonly Span<T> row4;
-        private readonly Span<T> row5;
-        private readonly Span<T> row6;
-        private readonly Span<T> row7;
-
-        public RowOctet(Buffer2D<T> buffer, int startY)
-        {
-            int y = startY;
-            int height = buffer.Height;
-            this.row0 = y < height ? buffer.GetRowSpan(y++) : default;
-            this.row1 = y < height ? buffer.GetRowSpan(y++) : default;
-            this.row2 = y < height ? buffer.GetRowSpan(y++) : default;
-            this.row3 = y < height ? buffer.GetRowSpan(y++) : default;
-            this.row4 = y < height ? buffer.GetRowSpan(y++) : default;
-            this.row5 = y < height ? buffer.GetRowSpan(y++) : default;
-            this.row6 = y < height ? buffer.GetRowSpan(y++) : default;
-            this.row7 = y < height ? buffer.GetRowSpan(y) : default;
-        }
+        private Span<T> row0;
+        private Span<T> row1;
+        private Span<T> row2;
+        private Span<T> row3;
+        private Span<T> row4;
+        private Span<T> row5;
+        private Span<T> row6;
+        private Span<T> row7;
 
+        // No unsafe tricks, since Span<T> can't be used as a generic argument
         public Span<T> this[int y]
         {
-            [MethodImpl(InliningOptions.ShortMethod)]
-            get
-            {
-                // No unsafe tricks, since Span<T> can't be used as a generic argument
-                return y switch
+            [MethodImpl(MethodImplOptions.AggressiveInlining)]
+            get =>
+                y switch
                 {
                     0 => this.row0,
                     1 => this.row1,
@@ -56,13 +41,57 @@ public Span<T> this[int y]
                     7 => this.row7,
                     _ => ThrowIndexOutOfRangeException()
                 };
+
+            [MethodImpl(MethodImplOptions.AggressiveInlining)]
+            private set
+            {
+                switch (y)
+                {
+                    case 0:
+                        this.row0 = value;
+                        break;
+                    case 1:
+                        this.row1 = value;
+                        break;
+                    case 2:
+                        this.row2 = value;
+                        break;
+                    case 3:
+                        this.row3 = value;
+                        break;
+                    case 4:
+                        this.row4 = value;
+                        break;
+                    case 5:
+                        this.row5 = value;
+                        break;
+                    case 6:
+                        this.row6 = value;
+                        break;
+                    default:
+                        this.row7 = value;
+                        break;
+                }
             }
         }
 
-        [MethodImpl(InliningOptions.ColdPath)]
-        private static Span<T> ThrowIndexOutOfRangeException()
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void Update(Buffer2D<T> buffer, int startY)
         {
-            throw new IndexOutOfRangeException();
+            // We don't actually have to assign values outside of the
+            // frame pixel buffer since they are never requested.
+            int y = startY;
+            int yEnd = Math.Min(y + 8, buffer.Height);
+
+            int i = 0;
+            while (y < yEnd)
+            {
+                this[i++] = buffer.GetRowSpan(y++);
+            }
         }
+
+        [MethodImpl(MethodImplOptions.NoInlining)]
+        private static Span<T> ThrowIndexOutOfRangeException()
+        => throw new IndexOutOfRangeException();
     }
 }