Add 4 channel float shuffling.

src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs

@@ -18,6 +18,86 @@ namespace SixLabors.ImageSharp
 
             public static ReadOnlySpan<byte> PermuteMaskEvenOdd8x32 => new byte[] { 0, 0, 0, 0, 2, 0, 0, 0, 4, 0, 0, 0, 6, 0, 0, 0, 1, 0, 0, 0, 3, 0, 0, 0, 5, 0, 0, 0, 7, 0, 0, 0 };
 
+            /// <summary>
+            /// Shuffle single-precision (32-bit) floating-point elements in <paramref name="source"/>
+            /// using the control and store the results in <paramref name="dest"/>.
+            /// </summary>
+            /// <param name="source">The source span of floats</param>
+            /// <param name="dest">The destination span of float</param>
+            /// <param name="control">The byte control.</param>
+            [MethodImpl(InliningOptions.ShortMethod)]
+            public static void Shuffle4ChannelReduce(
+                ref ReadOnlySpan<float> source,
+                ref Span<float> dest,
+                byte control)
+            {
+                if (Avx.IsSupported || Sse.IsSupported)
+                {
+                    int remainder;
+                    if (Avx.IsSupported)
+                    {
+                        remainder = ImageMaths.ModuloP2(source.Length, Vector256<float>.Count);
+                    }
+                    else
+                    {
+                        remainder = ImageMaths.ModuloP2(source.Length, Vector128<float>.Count);
+                    }
+
+                    int adjustedCount = source.Length - remainder;
+
+                    if (adjustedCount > 0)
+                    {
+                        Shuffle4Channel(
+                            source.Slice(0, adjustedCount),
+                            dest.Slice(0, adjustedCount),
+                            control);
+
+                        source = source.Slice(adjustedCount);
+                        dest = dest.Slice(adjustedCount);
+                    }
+                }
+            }
+
+            [MethodImpl(InliningOptions.ShortMethod)]
+            private static void Shuffle4Channel(
+                ReadOnlySpan<float> source,
+                Span<float> dest,
+                byte control)
+            {
+                if (Avx.IsSupported)
+                {
+                    int n = dest.Length / Vector256<float>.Count;
+
+                    ref Vector256<float> sourceBase =
+                        ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(source));
+
+                    ref Vector256<float> destBase =
+                        ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(dest));
+
+                    for (int i = 0; i < n; i++)
+                    {
+                        Unsafe.Add(ref destBase, i) = Avx.Permute(Unsafe.Add(ref sourceBase, i), control);
+                    }
+                }
+                else
+                {
+                    // Sse
+                    int n = dest.Length / Vector128<float>.Count;
+
+                    ref Vector128<float> sourceBase =
+                        ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(source));
+
+                    ref Vector128<float> destBase =
+                        ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(dest));
+
+                    for (int i = 0; i < n; i++)
+                    {
+                        Vector128<float> vs = Unsafe.Add(ref sourceBase, i);
+                        Unsafe.Add(ref destBase, i) = Sse.Shuffle(vs, vs, control);
+                    }
+                }
+            }
+
             /// <summary>
             /// Performs a multiplication and an addition of the <see cref="Vector256{T}"/>.
             /// </summary>

src/ImageSharp/Common/Helpers/SimdUtils.Shuffle.cs

@@ -0,0 +1,131 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
+using System;
+using System.Diagnostics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+
+namespace SixLabors.ImageSharp
+{
+    internal static partial class SimdUtils
+    {
+        /// <summary>
+        /// Shuffle single-precision (32-bit) floating-point elements in <paramref name="source"/>
+        /// using the control and store the results in <paramref name="dest"/>.
+        /// </summary>
+        /// <param name="source">The source span of floats</param>
+        /// <param name="dest">The destination span of float</param>
+        /// <param name="control">The byte control.</param>
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public static void Shuffle4Channel(
+            ReadOnlySpan<float> source,
+            Span<float> dest,
+            byte control)
+        {
+            VerifyShuffleSpanInput(source, dest);
+
+            // TODO: There doesn't seem to be any APIs for
+            // System.Numerics that allow shuffling.
+#if SUPPORTS_RUNTIME_INTRINSICS
+            HwIntrinsics.Shuffle4ChannelReduce(ref source, ref dest, control);
+#endif
+
+            // Deal with the remainder:
+            if (source.Length > 0)
+            {
+                ShuffleRemainder4Channel(source, dest, control);
+            }
+        }
+
+        [MethodImpl(InliningOptions.ColdPath)]
+        public static void ShuffleRemainder4Channel(
+            ReadOnlySpan<float> source,
+            Span<float> dest,
+            byte control)
+        {
+            ref float sBase = ref MemoryMarshal.GetReference(source);
+            ref float dBase = ref MemoryMarshal.GetReference(dest);
+            Shuffle.InverseMmShuffle(control, out int p3, out int p2, out int p1, out int p0);
+
+            for (int i = 0; i < source.Length; i += 4)
+            {
+                Unsafe.Add(ref dBase, i) = Unsafe.Add(ref sBase, p0 + i);
+                Unsafe.Add(ref dBase, i + 1) = Unsafe.Add(ref sBase, p1 + i);
+                Unsafe.Add(ref dBase, i + 2) = Unsafe.Add(ref sBase, p2 + i);
+                Unsafe.Add(ref dBase, i + 3) = Unsafe.Add(ref sBase, p3 + i);
+            }
+        }
+
+        [Conditional("DEBUG")]
+        private static void VerifyShuffleSpanInput(ReadOnlySpan<float> source, Span<float> dest)
+        {
+            DebugGuard.IsTrue(
+                source.Length == dest.Length,
+                nameof(source),
+                "Input spans must be of same length!");
+
+            DebugGuard.IsTrue(
+                source.Length % 4 == 0,
+                nameof(source),
+                "Input spans must be divisiable by 4!");
+        }
+
+        public static class Shuffle
+        {
+            public const byte WXYZ = (2 << 6) | (1 << 4) | (0 << 2) | 3;
+            public const byte XYZW = (3 << 6) | (2 << 4) | (1 << 2) | 0;
+            public const byte ZYXW = (3 << 6) | (0 << 4) | (1 << 2) | 2;
+
+            public static ReadOnlySpan<byte> WXYZ_128 => MmShuffleByte128(2, 1, 0, 3);
+
+            public static ReadOnlySpan<byte> XYZW_128 => MmShuffleByte128(3, 2, 1, 0);
+
+            public static ReadOnlySpan<byte> ZYXW_128 => MmShuffleByte128(3, 0, 1, 2);
+
+            public static ReadOnlySpan<byte> WXYZ_256 => MmShuffleByte256(2, 1, 0, 3);
+
+            public static ReadOnlySpan<byte> XYZW_256 => MmShuffleByte256(3, 2, 1, 0);
+
+            public static ReadOnlySpan<byte> ZYXW_256 => MmShuffleByte256(3, 0, 1, 2);
+
+            private static byte[] MmShuffleByte128(int p3, int p2, int p1, int p0)
+            {
+                byte[] result = new byte[16];
+
+                for (int i = 0; i < result.Length; i += 4)
+                {
+                    result[i] = (byte)(p0 + i);
+                    result[i + 1] = (byte)(p1 + i);
+                    result[i + 2] = (byte)(p2 + i);
+                    result[i + 3] = (byte)(p3 + i);
+                }
+
+                return result;
+            }
+
+            private static byte[] MmShuffleByte256(int p3, int p2, int p1, int p0)
+            {
+                byte[] result = new byte[32];
+
+                for (int i = 0; i < result.Length; i += 4)
+                {
+                    result[i] = (byte)(p0 + i);
+                    result[i + 1] = (byte)(p1 + i);
+                    result[i + 2] = (byte)(p2 + i);
+                    result[i + 3] = (byte)(p3 + i);
+                }
+
+                return result;
+            }
+
+            public static void InverseMmShuffle(byte control, out int p3, out int p2, out int p1, out int p0)
+            {
+                p3 = control >> 6 & 0x3;
+                p2 = control >> 4 & 0x3;
+                p1 = control >> 2 & 0x3;
+                p0 = control >> 0 & 0x3;
+            }
+        }
+    }
+}

tests/ImageSharp.Benchmarks/Color/Bulk/ShuffleFloat4Channel.cs

@@ -0,0 +1,68 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
+using System;
+using BenchmarkDotNet.Attributes;
+using SixLabors.ImageSharp.Tests;
+
+namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk
+{
+    [Config(typeof(Config.HwIntrinsics_SSE_AVX))]
+    public class ShuffleFloat4Channel
+    {
+        private float[] source;
+        private float[] destination;
+
+        [GlobalSetup]
+        public void Setup()
+        {
+            this.source = new Random(this.Count).GenerateRandomFloatArray(this.Count, 0, 256);
+            this.destination = new float[this.Count];
+        }
+
+        [Params(128, 256, 512, 1024, 2048)]
+        public int Count { get; set; }
+
+        [Benchmark]
+        public void Shuffle4Channel()
+        {
+            SimdUtils.Shuffle4Channel(this.source, this.destination, SimdUtils.Shuffle.WXYZ);
+        }
+    }
+
+    // 2020-10-26
+    // ##########
+    //
+    // BenchmarkDotNet=v0.12.1, OS=Windows 10.0.19041.572 (2004/?/20H1)
+    // Intel Core i7-8650U CPU 1.90GHz(Kaby Lake R), 1 CPU, 8 logical and 4 physical cores
+    // .NET Core SDK = 5.0.100-rc.2.20479.15
+    //
+    // [Host]          : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
+    //  AVX             : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
+    //  No HwIntrinsics : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
+    //  SSE             : .NET Core 3.1.9 (CoreCLR 4.700.20.47201, CoreFX 4.700.20.47203), X64 RyuJIT
+    //
+    // Runtime=.NET Core 3.1
+    //
+    // |          Method |             Job |                              EnvironmentVariables | Count |        Mean |    Error |   StdDev | Ratio | RatioSD | Gen 0 | Gen 1 | Gen 2 | Allocated |
+    // |---------------- |---------------- |-------------------------------------------------- |------ |------------:|---------:|---------:|------:|--------:|------:|------:|------:|----------:|
+    // | Shuffle4Channel |             AVX |                                             Empty |   128 |    14.49 ns | 0.244 ns | 0.217 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+    // | Shuffle4Channel | No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   128 |    87.74 ns | 0.524 ns | 0.490 ns |  6.06 |    0.09 |     - |     - |     - |         - |
+    // | Shuffle4Channel |             SSE |                               COMPlus_EnableAVX=0 |   128 |    23.65 ns | 0.101 ns | 0.094 ns |  1.63 |    0.03 |     - |     - |     - |         - |
+    // |                 |                 |                                                   |       |             |          |          |       |         |       |       |       |           |
+    // | Shuffle4Channel |             AVX |                                             Empty |   256 |    25.87 ns | 0.492 ns | 0.673 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+    // | Shuffle4Channel | No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   256 |   159.52 ns | 0.901 ns | 0.843 ns |  6.12 |    0.12 |     - |     - |     - |         - |
+    // | Shuffle4Channel |             SSE |                               COMPlus_EnableAVX=0 |   256 |    45.47 ns | 0.404 ns | 0.378 ns |  1.75 |    0.03 |     - |     - |     - |         - |
+    // |                 |                 |                                                   |       |             |          |          |       |         |       |       |       |           |
+    // | Shuffle4Channel |             AVX |                                             Empty |   512 |    49.51 ns | 0.088 ns | 0.083 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+    // | Shuffle4Channel | No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |   512 |   297.96 ns | 0.926 ns | 0.821 ns |  6.02 |    0.02 |     - |     - |     - |         - |
+    // | Shuffle4Channel |             SSE |                               COMPlus_EnableAVX=0 |   512 |    90.77 ns | 0.191 ns | 0.169 ns |  1.83 |    0.00 |     - |     - |     - |         - |
+    // |                 |                 |                                                   |       |             |          |          |       |         |       |       |       |           |
+    // | Shuffle4Channel |             AVX |                                             Empty |  1024 |   113.09 ns | 1.913 ns | 3.090 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+    // | Shuffle4Channel | No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  1024 |   604.58 ns | 1.464 ns | 1.298 ns |  5.29 |    0.18 |     - |     - |     - |         - |
+    // | Shuffle4Channel |             SSE |                               COMPlus_EnableAVX=0 |  1024 |   179.44 ns | 0.208 ns | 0.184 ns |  1.57 |    0.05 |     - |     - |     - |         - |
+    // |                 |                 |                                                   |       |             |          |          |       |         |       |       |       |           |
+    // | Shuffle4Channel |             AVX |                                             Empty |  2048 |   217.95 ns | 1.314 ns | 1.165 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+    // | Shuffle4Channel | No HwIntrinsics | COMPlus_EnableHWIntrinsic=0,COMPlus_FeatureSIMD=0 |  2048 | 1,152.04 ns | 3.941 ns | 3.494 ns |  5.29 |    0.03 |     - |     - |     - |         - |
+    // | Shuffle4Channel |             SSE |                               COMPlus_EnableAVX=0 |  2048 |   349.52 ns | 0.587 ns | 0.520 ns |  1.60 |    0.01 |     - |     - |     - |         - |
+}

tests/ImageSharp.Benchmarks/ImageSharp.Benchmarks.csproj

@@ -17,6 +17,7 @@
   <ItemGroup>
     <Compile Include="..\ImageSharp.Tests\TestImages.cs" Link="Tests\TestImages.cs" />
     <Compile Include="..\ImageSharp.Tests\TestUtilities\TestEnvironment.cs" Link="Tests\TestEnvironment.cs" />
+    <Compile Include="..\ImageSharp.Tests\TestUtilities\TestDataGenerator.cs" Link="Tests\TestDataGenerator.cs" />
   </ItemGroup>
 
   <ItemGroup>

tests/ImageSharp.Tests/Common/SimdUtilsTests.Shuffle.cs

@@ -0,0 +1,75 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
+using System;
+using SixLabors.ImageSharp.Tests.TestUtilities;
+using Xunit;
+
+namespace SixLabors.ImageSharp.Tests.Common
+{
+    public partial class SimdUtilsTests
+    {
+        public static readonly TheoryData<byte> ShuffleControls =
+            new TheoryData<byte>
+            {
+                SimdUtils.Shuffle.WXYZ,
+                SimdUtils.Shuffle.XYZW,
+                SimdUtils.Shuffle.ZYXW
+            };
+
+        [Theory]
+        [MemberData(nameof(ShuffleControls))]
+        public void BulkShuffleFloat4Channel(byte control)
+        {
+            static void RunTest(string serialized)
+            {
+                byte ctrl = FeatureTestRunner.Deserialize<byte>(serialized);
+                foreach (var item in ArraySizesDivisibleBy4)
+                {
+                    foreach (var count in item)
+                    {
+                        TestShuffle(
+                            (int)count,
+                            (s, d) => SimdUtils.Shuffle4Channel(s.Span, d.Span, ctrl),
+                            ctrl);
+                    }
+                }
+            }
+
+            FeatureTestRunner.RunWithHwIntrinsicsFeature(
+                RunTest,
+                control,
+                HwIntrinsics.AllowAll | HwIntrinsics.DisableAVX2 | HwIntrinsics.DisableSSE);
+        }
+
+        private static void TestShuffle(
+            int count,
+            Action<Memory<float>, Memory<float>> convert,
+            byte control)
+        {
+            float[] source = new Random(count).GenerateRandomFloatArray(count, 0, 256);
+            var result = new float[count];
+
+            float[] expected = new float[count];
+
+            SimdUtils.Shuffle.InverseMmShuffle(
+                control,
+                out int p3,
+                out int p2,
+                out int p1,
+                out int p0);
+
+            for (int i = 0; i < expected.Length; i += 4)
+            {
+                expected[i] = source[p0 + i];
+                expected[i + 1] = source[p1 + i];
+                expected[i + 2] = source[p2 + i];
+                expected[i + 3] = source[p3 + i];
+            }
+
+            convert(source, result);
+
+            Assert.Equal(expected, result, new ApproximateFloatComparer(1e-5F));
+        }
+    }
+}

tests/ImageSharp.Tests/Common/SimdUtilsTests.cs

@@ -13,7 +13,7 @@ using Xunit.Abstractions;
 
 namespace SixLabors.ImageSharp.Tests.Common
 {
-    public class SimdUtilsTests
+    public partial class SimdUtilsTests
     {
         private ITestOutputHelper Output { get; }
 
@@ -212,14 +212,14 @@ namespace SixLabors.ImageSharp.Tests.Common
             static void RunTest(string serialized)
             {
                 TestImpl_BulkConvertByteToNormalizedFloat(
-                    FeatureTestRunner.Deserialize(serialized),
+                    FeatureTestRunner.Deserialize<int>(serialized),
                     (s, d) => SimdUtils.HwIntrinsics.ByteToNormalizedFloat(s.Span, d.Span));
             }
 
             FeatureTestRunner.RunWithHwIntrinsicsFeature(
                 RunTest,
-                HwIntrinsics.AllowAll | HwIntrinsics.DisableAVX2 | HwIntrinsics.DisableSSE41,
-                count);
+                count,
+                HwIntrinsics.AllowAll | HwIntrinsics.DisableAVX2 | HwIntrinsics.DisableSSE41);
         }
 #endif
 
@@ -305,14 +305,14 @@ namespace SixLabors.ImageSharp.Tests.Common
             static void RunTest(string serialized)
             {
                 TestImpl_BulkConvertNormalizedFloatToByteClampOverflows(
-                    FeatureTestRunner.Deserialize(serialized),
+                    FeatureTestRunner.Deserialize<int>(serialized),
                     (s, d) => SimdUtils.HwIntrinsics.NormalizedFloatToByteSaturate(s.Span, d.Span));
             }
 
             FeatureTestRunner.RunWithHwIntrinsicsFeature(
                 RunTest,
-                HwIntrinsics.AllowAll | HwIntrinsics.DisableAVX2,
-                count);
+                count,
+                HwIntrinsics.AllowAll | HwIntrinsics.DisableAVX2);
         }
 
 #endif

tests/ImageSharp.Tests/Formats/Png/PngEncoderTests.cs

@@ -535,7 +535,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Png
             static void RunTest(string serialized)
             {
                 TestImageProvider<Rgba32> provider =
-                    FeatureTestRunner.Deserialize<TestImageProvider<Rgba32>>(serialized);
+                    FeatureTestRunner.DeserializeForXunit<TestImageProvider<Rgba32>>(serialized);
 
                 foreach (PngInterlaceMode interlaceMode in InterlaceMode)
                 {

tests/ImageSharp.Tests/TestUtilities/FeatureTesting/FeatureTestRunner.cs

@@ -29,17 +29,19 @@ namespace SixLabors.ImageSharp.Tests.TestUtilities
         /// <typeparam name="T">The type to deserialize to.</typeparam>
         /// <param name="value">The string value to deserialize.</param>
         /// <returns>The <see cref="T"/> value.</returns>
-        public static T Deserialize<T>(string value)
+        public static T DeserializeForXunit<T>(string value)
             where T : IXunitSerializable
             => BasicSerializer.Deserialize<T>(value);
 
         /// <summary>
-        /// Allows the deserialization of integers passed to the feature test.
+        /// Allows the deserialization of types implementing <see cref="IConvertible"/>
+        /// passed to the feature test.
         /// </summary>
         /// <param name="value">The string value to deserialize.</param>
-        /// <returns>The <see cref="int"/> value.</returns>
-        public static int Deserialize(string value)
-            => Convert.ToInt32(value);
+        /// <returns>The <typeparamref name="T"/> value.</returns>
+        public static T Deserialize<T>(string value)
+            where T : IConvertible
+            => (T)Convert.ChangeType(value, typeof(T));
 
         /// <summary>
         /// Runs the given test <paramref name="action"/> within an environment
@@ -214,12 +216,13 @@ namespace SixLabors.ImageSharp.Tests.TestUtilities
         /// where the given <paramref name="intrinsics"/> features.
         /// </summary>
         /// <param name="action">The test action to run.</param>
-        /// <param name="intrinsics">The intrinsics features.</param>
         /// <param name="serializable">The value to pass as a parameter to the test action.</param>
-        public static void RunWithHwIntrinsicsFeature(
+        /// <param name="intrinsics">The intrinsics features.</param>
+        public static void RunWithHwIntrinsicsFeature<T>(
             Action<string> action,
-            HwIntrinsics intrinsics,
-            int serializable)
+            T serializable,
+            HwIntrinsics intrinsics)
+            where T : IConvertible
         {
             if (!RemoteExecutor.IsSupported)
             {

tests/ImageSharp.Tests/TestUtilities/Tests/FeatureTestRunnerTests.cs

@@ -183,7 +183,7 @@ namespace SixLabors.ImageSharp.Tests.TestUtilities.Tests
             static void AssertHwIntrinsicsFeatureDisabled(string serializable)
             {
                 Assert.NotNull(serializable);
-                Assert.NotNull(FeatureTestRunner.Deserialize<FakeSerializable>(serializable));
+                Assert.NotNull(FeatureTestRunner.DeserializeForXunit<FakeSerializable>(serializable));
 
 #if SUPPORTS_RUNTIME_INTRINSICS
                 Assert.False(Sse.IsSupported);
@@ -202,7 +202,7 @@ namespace SixLabors.ImageSharp.Tests.TestUtilities.Tests
             static void AssertHwIntrinsicsFeatureDisabled(string serializable, string intrinsic)
             {
                 Assert.NotNull(serializable);
-                Assert.NotNull(FeatureTestRunner.Deserialize<FakeSerializable>(serializable));
+                Assert.NotNull(FeatureTestRunner.DeserializeForXunit<FakeSerializable>(serializable));
 
                 switch ((HwIntrinsics)Enum.Parse(typeof(HwIntrinsics), intrinsic))
                 {