Merge pull request #1481 from SixLabors/js/faster-resize

Much Faster sRGB Companding

Author: JimBobSquarePants

src/ImageSharp/ColorSpaces/Companding/SRgbCompanding.cs

@@ -5,6 +5,10 @@
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
+#if SUPPORTS_RUNTIME_INTRINSICS
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+#endif
 
 namespace SixLabors.ImageSharp.ColorSpaces.Companding
 {
@@ -18,49 +22,119 @@ namespace SixLabors.ImageSharp.ColorSpaces.Companding
     /// </remarks>
     public static class SRgbCompanding
     {
+        private const int Length = Scale + 2; // 256kb @ 16bit precision.
+        private const int Scale = (1 << 16) - 1;
+
+        private static readonly Lazy<float[]> LazyCompressTable = new Lazy<float[]>(
+            () =>
+            {
+                var result = new float[Length];
+
+                for (int i = 0; i < result.Length; i++)
+                {
+                    double d = (double)i / Scale;
+                    if (d <= (0.04045 / 12.92))
+                    {
+                        d *= 12.92;
+                    }
+                    else
+                    {
+                        d = (1.055 * Math.Pow(d, 1.0 / 2.4)) - 0.055;
+                    }
+
+                    result[i] = (float)d;
+                }
+
+                return result;
+            },
+            true);
+
+        private static readonly Lazy<float[]> LazyExpandTable = new Lazy<float[]>(
+            () =>
+            {
+                var result = new float[Length];
+
+                for (int i = 0; i < result.Length; i++)
+                {
+                    double d = (double)i / Scale;
+                    if (d <= 0.04045)
+                    {
+                        d /= 12.92;
+                    }
+                    else
+                    {
+                        d = Math.Pow((d + 0.055) / 1.055, 2.4);
+                    }
+
+                    result[i] = (float)d;
+                }
+
+                return result;
+            },
+            true);
+
+        private static float[] ExpandTable => LazyExpandTable.Value;
+
+        private static float[] CompressTable => LazyCompressTable.Value;
+
         /// <summary>
         /// Expands the companded vectors to their linear equivalents with respect to the energy.
         /// </summary>
         /// <param name="vectors">The span of vectors.</param>
-        [MethodImpl(InliningOptions.ShortMethod)]
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static void Expand(Span<Vector4> vectors)
         {
-            ref Vector4 vectorsStart = ref MemoryMarshal.GetReference(vectors);
-            ref Vector4 vectorsEnd = ref Unsafe.Add(ref vectorsStart, vectors.Length);
-
-            while (Unsafe.IsAddressLessThan(ref vectorsStart, ref vectorsEnd))
+#if SUPPORTS_RUNTIME_INTRINSICS
+            if (Avx2.IsSupported && vectors.Length >= 2)
             {
-                Expand(ref vectorsStart);
+                CompandAvx2(vectors, ExpandTable);
 
-                vectorsStart = ref Unsafe.Add(ref vectorsStart, 1);
+                if (Numerics.Modulo2(vectors.Length) != 0)
+                {
+                    // Vector4 fits neatly in pairs. Any overlap has to be equal to 1.
+                    Expand(ref MemoryMarshal.GetReference(vectors.Slice(vectors.Length - 1)));
+                }
+            }
+            else
+#endif
+            {
+                CompandScalar(vectors, ExpandTable);
             }
         }
 
         /// <summary>
         /// Compresses the uncompanded vectors to their nonlinear equivalents with respect to the energy.
         /// </summary>
         /// <param name="vectors">The span of vectors.</param>
-        [MethodImpl(InliningOptions.ShortMethod)]
-        public static void Compress(Span<Vector4> vectors)
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static unsafe void Compress(Span<Vector4> vectors)
         {
-            ref Vector4 vectorsStart = ref MemoryMarshal.GetReference(vectors);
-            ref Vector4 vectorsEnd = ref Unsafe.Add(ref vectorsStart, vectors.Length);
-
-            while (Unsafe.IsAddressLessThan(ref vectorsStart, ref vectorsEnd))
+#if SUPPORTS_RUNTIME_INTRINSICS
+            if (Avx2.IsSupported && vectors.Length >= 2)
             {
-                Compress(ref vectorsStart);
+                CompandAvx2(vectors, CompressTable);
 
-                vectorsStart = ref Unsafe.Add(ref vectorsStart, 1);
+                if (Numerics.Modulo2(vectors.Length) != 0)
+                {
+                    // Vector4 fits neatly in pairs. Any overlap has to be equal to 1.
+                    Compress(ref MemoryMarshal.GetReference(vectors.Slice(vectors.Length - 1)));
+                }
+            }
+            else
+#endif
+            {
+                CompandScalar(vectors, CompressTable);
             }
         }
 
         /// <summary>
         /// Expands a companded vector to its linear equivalent with respect to the energy.
         /// </summary>
         /// <param name="vector">The vector.</param>
-        [MethodImpl(InliningOptions.ShortMethod)]
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static void Expand(ref Vector4 vector)
         {
+            // Alpha is already a linear representation of opacity so we do not want to convert it.
             vector.X = Expand(vector.X);
             vector.Y = Expand(vector.Y);
             vector.Z = Expand(vector.Z);
@@ -70,9 +144,10 @@ public static void Expand(ref Vector4 vector)
         /// Compresses an uncompanded vector (linear) to its nonlinear equivalent.
         /// </summary>
         /// <param name="vector">The vector.</param>
-        [MethodImpl(InliningOptions.ShortMethod)]
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public static void Compress(ref Vector4 vector)
         {
+            // Alpha is already a linear representation of opacity so we do not want to convert it.
             vector.X = Compress(vector.X);
             vector.Y = Compress(vector.Y);
             vector.Z = Compress(vector.Z);
@@ -83,15 +158,84 @@ public static void Compress(ref Vector4 vector)
         /// </summary>
         /// <param name="channel">The channel value.</param>
         /// <returns>The <see cref="float"/> representing the linear channel value.</returns>
-        [MethodImpl(InliningOptions.ShortMethod)]
-        public static float Expand(float channel) => channel <= 0.04045F ? channel / 12.92F : MathF.Pow((channel + 0.055F) / 1.055F, 2.4F);
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static float Expand(float channel)
+            => channel <= 0.04045F ? channel / 12.92F : MathF.Pow((channel + 0.055F) / 1.055F, 2.4F);
 
         /// <summary>
         /// Compresses an uncompanded channel (linear) to its nonlinear equivalent.
         /// </summary>
         /// <param name="channel">The channel value.</param>
         /// <returns>The <see cref="float"/> representing the nonlinear channel value.</returns>
-        [MethodImpl(InliningOptions.ShortMethod)]
-        public static float Compress(float channel) => channel <= 0.0031308F ? 12.92F * channel : (1.055F * MathF.Pow(channel, 0.416666666666667F)) - 0.055F;
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static float Compress(float channel)
+            => channel <= 0.0031308F ? 12.92F * channel : (1.055F * MathF.Pow(channel, 0.416666666666667F)) - 0.055F;
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static unsafe void CompandAvx2(Span<Vector4> vectors, float[] table)
+        {
+            fixed (float* tablePointer = &table[0])
+            {
+                var scale = Vector256.Create((float)Scale);
+                Vector256<float> zero = Vector256<float>.Zero;
+                var offset = Vector256.Create(1);
+
+                // Divide by 2 as 4 elements per Vector4 and 8 per Vector256<float>
+                ref Vector256<float> vectorsBase = ref Unsafe.As<Vector4, Vector256<float>>(ref MemoryMarshal.GetReference(vectors));
+                ref Vector256<float> vectorsLast = ref Unsafe.Add(ref vectorsBase, (IntPtr)((uint)vectors.Length / 2u));
+
+                while (Unsafe.IsAddressLessThan(ref vectorsBase, ref vectorsLast))
+                {
+                    Vector256<float> multiplied = Avx.Multiply(scale, vectorsBase);
+                    multiplied = Avx.Min(Avx.Max(zero, multiplied), scale);
+
+                    Vector256<int> truncated = Avx.ConvertToVector256Int32WithTruncation(multiplied);
+                    Vector256<float> truncatedF = Avx.ConvertToVector256Single(truncated);
+
+                    Vector256<float> low = Avx2.GatherVector256(tablePointer, truncated, sizeof(float));
+                    Vector256<float> high = Avx2.GatherVector256(tablePointer, Avx2.Add(truncated, offset), sizeof(float));
+
+                    // Alpha is already a linear representation of opacity so we do not want to convert it.
+                    Vector256<float> companded = Numerics.Lerp(low, high, Avx.Subtract(multiplied, truncatedF));
+                    vectorsBase = Avx.Blend(companded, vectorsBase, Numerics.BlendAlphaControl);
+                    vectorsBase = ref Unsafe.Add(ref vectorsBase, 1);
+                }
+            }
+        }
+#endif
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static unsafe void CompandScalar(Span<Vector4> vectors, float[] table)
+        {
+            fixed (float* tablePointer = &table[0])
+            {
+                Vector4 zero = Vector4.Zero;
+                var scale = new Vector4(Scale);
+                ref Vector4 vectorsBase = ref MemoryMarshal.GetReference(vectors);
+                ref Vector4 vectorsLast = ref Unsafe.Add(ref vectorsBase, vectors.Length);
+
+                while (Unsafe.IsAddressLessThan(ref vectorsBase, ref vectorsLast))
+                {
+                    Vector4 multiplied = Numerics.Clamp(vectorsBase * Scale, zero, scale);
+
+                    float f0 = multiplied.X;
+                    float f1 = multiplied.Y;
+                    float f2 = multiplied.Z;
+
+                    uint i0 = (uint)f0;
+                    uint i1 = (uint)f1;
+                    uint i2 = (uint)f2;
+
+                    // Alpha is already a linear representation of opacity so we do not want to convert it.
+                    vectorsBase.X = Numerics.Lerp(tablePointer[i0], tablePointer[i0 + 1], f0 - (int)i0);
+                    vectorsBase.Y = Numerics.Lerp(tablePointer[i1], tablePointer[i1 + 1], f1 - (int)i1);
+                    vectorsBase.Z = Numerics.Lerp(tablePointer[i2], tablePointer[i2 + 1], f2 - (int)i2);
+
+                    vectorsBase = ref Unsafe.Add(ref vectorsBase, 1);
+                }
+            }
+        }
     }
 }

src/ImageSharp/Common/Helpers/Numerics.cs

@@ -19,7 +19,7 @@ namespace SixLabors.ImageSharp
     internal static class Numerics
     {
 #if SUPPORTS_RUNTIME_INTRINSICS
-        private const int BlendAlphaControl = 0b_10_00_10_00;
+        public const int BlendAlphaControl = 0b_10_00_10_00;
         private const int ShuffleAlphaControl = 0b_11_11_11_11;
 #endif
 
@@ -710,5 +710,43 @@ public static unsafe void CubeRootOnXYZ(Span<Vector4> vectors)
                 }
             }
         }
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+
+        /// <summary>
+        /// Performs a linear interpolation between two values based on the given weighting.
+        /// </summary>
+        /// <param name="value1">The first value.</param>
+        /// <param name="value2">The second value.</param>
+        /// <param name="amount">Values between 0 and 1 that indicates the weight of <paramref name="value2"/>.</param>
+        /// <returns>The <see cref="Vector256{Single}"/>.</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static Vector256<float> Lerp(
+            in Vector256<float> value1,
+            in Vector256<float> value2,
+            in Vector256<float> amount)
+        {
+            Vector256<float> diff = Avx.Subtract(value2, value1);
+            if (Fma.IsSupported)
+            {
+                return Fma.MultiplyAdd(diff, amount, value1);
+            }
+            else
+            {
+                return Avx.Add(Avx.Multiply(diff, amount), value1);
+            }
+        }
+#endif
+
+        /// <summary>
+        /// Performs a linear interpolation between two values based on the given weighting.
+        /// </summary>
+        /// <param name="value1">The first value.</param>
+        /// <param name="value2">The second value.</param>
+        /// <param name="amount">A value between 0 and 1 that indicates the weight of <paramref name="value2"/>.</param>
+        /// <returns>The <see cref="float"/>.</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public static float Lerp(float value1, float value2, float amount)
+            => ((value2 - value1) * amount) + value1;
     }
 }

src/ImageSharp/PixelFormats/Utils/Vector4Converters.cs

@@ -1,4 +1,4 @@
-// Copyright (c) Six Labors.
+// Copyright (c) Six Labors.
 // Licensed under the Apache License, Version 2.0.
 
 using System;
@@ -45,4 +45,4 @@ internal static void ApplyBackwardConversionModifiers(Span<Vector4> vectors, Pix
             }
         }
     }
-}
+}

src/ImageSharp/Processing/Processors/Transforms/Resize/ResizeKernel.cs

@@ -61,9 +61,7 @@ public Span<float> Values
         /// <returns>The weighted sum</returns>
         [MethodImpl(InliningOptions.ShortMethod)]
         public Vector4 Convolve(Span<Vector4> rowSpan)
-        {
-            return this.ConvolveCore(ref rowSpan[this.StartIndex]);
-        }
+            => this.ConvolveCore(ref rowSpan[this.StartIndex]);
 
         [MethodImpl(InliningOptions.ShortMethod)]
         public Vector4 ConvolveCore(ref Vector4 rowStartRef)
@@ -91,9 +89,7 @@ public Vector4 ConvolveCore(ref Vector4 rowStartRef)
         /// </summary>
         [MethodImpl(InliningOptions.ShortMethod)]
         internal ResizeKernel AlterLeftValue(int left)
-        {
-            return new ResizeKernel(left, this.bufferPtr, this.Length);
-        }
+            => new ResizeKernel(left, this.bufferPtr, this.Length);
 
         internal void Fill(Span<double> values)
         {

src/ImageSharp/Processing/Processors/Transforms/Resize/ResizeWorker.cs

@@ -105,14 +105,10 @@ public void Dispose()
 
         [MethodImpl(InliningOptions.ShortMethod)]
         public Span<Vector4> GetColumnSpan(int x, int startY)
-        {
-            return this.transposedFirstPassBuffer.GetRowSpan(x).Slice(startY - this.currentWindow.Min);
-        }
+            => this.transposedFirstPassBuffer.GetRowSpan(x).Slice(startY - this.currentWindow.Min);
 
         public void Initialize()
-        {
-            this.CalculateFirstPassValues(this.currentWindow);
-        }
+            => this.CalculateFirstPassValues(this.currentWindow);
 
         public void FillDestinationPixels(RowInterval rowInterval, Buffer2D<TPixel> destination)
         {