polishing

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

@@ -30,7 +30,7 @@ namespace SixLabors.ImageSharp
                     0, 1, 2, 4, 5, 6, 8, 9, 10, 12, 13, 14, 3, 7, 11, 15
                 };
 
-            public static ReadOnlySpan<byte> PermuteMaskShiftAlpha8x32 => 
+            public static ReadOnlySpan<byte> PermuteMaskShiftAlpha8x32 =>
                 new byte[]
                 {
                     0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 4, 0, 0, 0,
@@ -877,6 +877,65 @@ namespace SixLabors.ImageSharp
                 blueChannel = blueChannel.Slice(slice);
                 destination = destination.Slice(slice);
             }
+
+            internal static void PackFromRgbPlanesAvx2Reduce(
+                ref ReadOnlySpan<byte> redChannel,
+                ref ReadOnlySpan<byte> greenChannel,
+                ref ReadOnlySpan<byte> blueChannel,
+                ref Span<Rgba32> destination)
+            {
+                ref Vector256<byte> rBase = ref Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(redChannel));
+                ref Vector256<byte> gBase = ref Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(greenChannel));
+                ref Vector256<byte> bBase = ref Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(blueChannel));
+                ref Vector256<byte> dBase = ref Unsafe.As<Rgba32, Vector256<byte>>(ref MemoryMarshal.GetReference(destination));
+
+                int count = redChannel.Length / Vector256<byte>.Count;
+
+                ref byte control1Bytes = ref MemoryMarshal.GetReference(SimdUtils.HwIntrinsics.PermuteMaskEvenOdd8x32);
+                Vector256<uint> control1 = Unsafe.As<byte, Vector256<uint>>(ref control1Bytes);
+
+                ref byte control2Bytes = ref MemoryMarshal.GetReference(PermuteMaskShiftAlpha8x32);
+                Vector256<uint> control2 = Unsafe.As<byte, Vector256<uint>>(ref control2Bytes);
+
+                Vector256<byte> a = Vector256.Create((byte)255);
+
+                Vector256<byte> shuffleAlpha = Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(ShuffleMaskShiftAlpha));
+
+                for (int i = 0; i < count; i++)
+                {
+                    Vector256<byte> r0 = Unsafe.Add(ref rBase, i);
+                    Vector256<byte> g0 = Unsafe.Add(ref gBase, i);
+                    Vector256<byte> b0 = Unsafe.Add(ref bBase, i);
+
+                    r0 = Avx2.PermuteVar8x32(r0.AsUInt32(), control1).AsByte();
+                    g0 = Avx2.PermuteVar8x32(g0.AsUInt32(), control1).AsByte();
+                    b0 = Avx2.PermuteVar8x32(b0.AsUInt32(), control1).AsByte();
+
+                    Vector256<byte> rg = Avx2.UnpackLow(r0, g0);
+                    Vector256<byte> b1 = Avx2.UnpackLow(b0, a);
+
+                    Vector256<byte> rgb1 = Avx2.UnpackLow(rg.AsUInt16(), b1.AsUInt16()).AsByte();
+                    Vector256<byte> rgb2 = Avx2.UnpackHigh(rg.AsUInt16(), b1.AsUInt16()).AsByte();
+
+                    rg = Avx2.UnpackHigh(r0, g0);
+                    b1 = Avx2.UnpackHigh(b0, a);
+
+                    Vector256<byte> rgb3 = Avx2.UnpackLow(rg.AsUInt16(), b1.AsUInt16()).AsByte();
+                    Vector256<byte> rgb4 = Avx2.UnpackHigh(rg.AsUInt16(), b1.AsUInt16()).AsByte();
+
+                    ref Vector256<byte> d0 = ref Unsafe.Add(ref dBase, i * 4);
+                    d0 = rgb1;
+                    Unsafe.Add(ref d0, 1) = rgb2;
+                    Unsafe.Add(ref d0, 2) = rgb3;
+                    Unsafe.Add(ref d0, 3) = rgb4;
+                }
+
+                int slice = count * Vector256<byte>.Count;
+                redChannel = redChannel.Slice(slice);
+                greenChannel = greenChannel.Slice(slice);
+                blueChannel = blueChannel.Slice(slice);
+                destination = destination.Slice(slice);
+            }
         }
     }
 }

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

@@ -1,3 +1,6 @@
+// Copyright (c) Six Labors.
+// Licensed under the Apache License, Version 2.0.
+
 using System;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
@@ -20,12 +23,9 @@ namespace SixLabors.ImageSharp
             ReadOnlySpan<byte> blueChannel,
             Span<Rgb24> destination)
         {
-            int count = redChannel.Length;
-            DebugGuard.IsTrue(greenChannel.Length == count, "Channels must be of same size!");
-            DebugGuard.IsTrue(blueChannel.Length == count, "Channels must be of same size!");
-
-            // To avoid overflows, this check is not debug-only:
-            Guard.IsTrue(destination.Length > count + 2, nameof(destination), "'destination' must contain a padding of 3 elements!");
+            DebugGuard.IsTrue(greenChannel.Length == redChannel.Length, nameof(greenChannel), "Channels must be of same size!");
+            DebugGuard.IsTrue(blueChannel.Length == redChannel.Length, nameof(blueChannel), "Channels must be of same size!");
+            DebugGuard.IsTrue(destination.Length > redChannel.Length + 2, nameof(destination), "'destination' must contain a padding of 3 elements!");
 
 #if SUPPORTS_RUNTIME_INTRINSICS
             if (Avx2.IsSupported)
@@ -49,7 +49,21 @@ namespace SixLabors.ImageSharp
             ReadOnlySpan<byte> blueChannel,
             Span<Rgba32> destination)
         {
-            PackFromRgbPlanesScalarBatchedReduce(ref redChannel, ref greenChannel, ref blueChannel, ref destination);
+            DebugGuard.IsTrue(greenChannel.Length == redChannel.Length, nameof(greenChannel), "Channels must be of same size!");
+            DebugGuard.IsTrue(blueChannel.Length == redChannel.Length, nameof(blueChannel), "Channels must be of same size!");
+            DebugGuard.IsTrue(destination.Length > redChannel.Length, nameof(destination), "'destination' span should not be shorter than the source channels!");
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+            if (Avx2.IsSupported)
+            {
+                HwIntrinsics.PackFromRgbPlanesAvx2Reduce(ref redChannel, ref greenChannel, ref blueChannel, ref destination);
+            }
+            else
+#endif
+            {
+                PackFromRgbPlanesScalarBatchedReduce(ref redChannel, ref greenChannel, ref blueChannel, ref destination);
+            }
+
             PackFromRgbPlanesRemainder(redChannel, greenChannel, blueChannel, destination);
         }
 

src/ImageSharp/PixelFormats/PixelImplementations/PixelOperations/Rgb24.PixelOperations.cs

@@ -21,6 +21,22 @@ namespace SixLabors.ImageSharp.PixelFormats
 
             /// <inheritdoc />
             public override PixelTypeInfo GetPixelTypeInfo() => LazyInfo.Value;
+
+            /// <inheritdoc />
+            internal override void PackFromRgbPlanes(
+                Configuration configuration,
+                ReadOnlySpan<byte> redChannel,
+                ReadOnlySpan<byte> greenChannel,
+                ReadOnlySpan<byte> blueChannel,
+                Span<Rgb24> destination)
+            {
+                int count = redChannel.Length;
+                Guard.IsTrue(greenChannel.Length == count, nameof(greenChannel), "Channels must be of same size!");
+                Guard.IsTrue(blueChannel.Length == count, nameof(blueChannel), "Channels must be of same size!");
+                Guard.IsTrue(destination.Length > count + 2, nameof(destination), "'destination' must contain a padding of 3 elements!");
+
+                SimdUtils.PackFromRgbPlanes(configuration, redChannel, greenChannel, blueChannel, destination);
+            }
         }
     }
 }

src/ImageSharp/PixelFormats/PixelImplementations/PixelOperations/Rgba32.PixelOperations.cs

@@ -56,6 +56,22 @@ namespace SixLabors.ImageSharp.PixelFormats
                     MemoryMarshal.Cast<Vector4, float>(sourceVectors),
                     MemoryMarshal.Cast<Rgba32, byte>(destinationPixels));
             }
+
+            /// <inheritdoc />
+            internal override void PackFromRgbPlanes(
+                Configuration configuration,
+                ReadOnlySpan<byte> redChannel,
+                ReadOnlySpan<byte> greenChannel,
+                ReadOnlySpan<byte> blueChannel,
+                Span<Rgba32> destination)
+            {
+                int count = redChannel.Length;
+                Guard.IsTrue(greenChannel.Length == count, nameof(greenChannel), "Channels must be of same size!");
+                Guard.IsTrue(blueChannel.Length == count, nameof(blueChannel), "Channels must be of same size!");
+                Guard.IsTrue(destination.Length > count, nameof(destination), "'destination' span should not be shorter than the source channels!");
+
+                SimdUtils.PackFromRgbPlanes(configuration, redChannel, greenChannel, blueChannel, destination);
+            }
         }
     }
 }

tests/ImageSharp.Benchmarks/General/PixelConversion/PixelConversion_PackFromRgbPlanes.cs

@@ -196,7 +196,7 @@ namespace SixLabors.ImageSharp.Benchmarks.General.PixelConversion
 
 #if SUPPORTS_RUNTIME_INTRINSICS
         [Benchmark(Baseline = true)]
-        public void Rgba32_Vector_Float()
+        public void Rgba32_Avx2_Float()
         {
             ref Vector256<float> rBase = ref Unsafe.As<float, Vector256<float>>(ref this.rFloat[0]);
             ref Vector256<float> gBase = ref Unsafe.As<float, Vector256<float>>(ref this.gFloat[0]);
@@ -237,7 +237,7 @@ namespace SixLabors.ImageSharp.Benchmarks.General.PixelConversion
         }
 
         [Benchmark]
-        public void Rgba32_Vector_Bytes()
+        public void Rgba24_Avx2_Bytes()
         {
             ReadOnlySpan<byte> r = this.rBuf;
             ReadOnlySpan<byte> g = this.rBuf;
@@ -245,6 +245,16 @@ namespace SixLabors.ImageSharp.Benchmarks.General.PixelConversion
             Span<Rgb24> rgb = this.rgbBuf;
             SimdUtils.HwIntrinsics.PackFromRgbPlanesAvx2Reduce(ref r, ref g, ref b, ref rgb);
         }
+
+        [Benchmark]
+        public void Rgba32_Avx2_Bytes()
+        {
+            ReadOnlySpan<byte> r = this.rBuf;
+            ReadOnlySpan<byte> g = this.rBuf;
+            ReadOnlySpan<byte> b = this.rBuf;
+            Span<Rgba32> rgb = this.rgbaBuf;
+            SimdUtils.HwIntrinsics.PackFromRgbPlanesAvx2Reduce(ref r, ref g, ref b, ref rgb);
+        }
 #endif
 
 #pragma warning disable SA1132

tests/ImageSharp.Tests/Common/SimdUtilsTests.cs

@@ -392,6 +392,41 @@ namespace SixLabors.ImageSharp.Tests.Common
             Assert.Equal(0, bb.Length);
             Assert.Equal(padding, dd.Length);
         }
+
+        [Fact]
+        public void PackFromRgbPlanesAvx2Reduce_Rgba32()
+        {
+            if (!Avx2.IsSupported)
+            {
+                return;
+            }
+
+            byte[] r = Enumerable.Range(0, 32).Select(x => (byte)x).ToArray();
+            byte[] g = Enumerable.Range(100, 32).Select(x => (byte)x).ToArray();
+            byte[] b = Enumerable.Range(200, 32).Select(x => (byte)x).ToArray();
+
+            Rgba32[] d = new Rgba32[32];
+
+            ReadOnlySpan<byte> rr = r.AsSpan();
+            ReadOnlySpan<byte> gg = g.AsSpan();
+            ReadOnlySpan<byte> bb = b.AsSpan();
+            Span<Rgba32> dd = d.AsSpan();
+
+            SimdUtils.HwIntrinsics.PackFromRgbPlanesAvx2Reduce(ref rr, ref gg, ref bb, ref dd);
+
+            for (int i = 0; i < 32; i++)
+            {
+                Assert.Equal(i, d[i].R);
+                Assert.Equal(i + 100, d[i].G);
+                Assert.Equal(i + 200, d[i].B);
+                Assert.Equal(255, d[i].A);
+            }
+
+            Assert.Equal(0, rr.Length);
+            Assert.Equal(0, gg.Length);
+            Assert.Equal(0, bb.Length);
+            Assert.Equal(0, dd.Length);
+        }
 #endif
 
         internal static void TestPackFromRgbPlanes<TPixel>(int count, Action<byte[], byte[], byte[], TPixel[]> packMethod)