Improved scalar transpose implementation

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

@@ -632,22 +632,55 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         [MethodImpl(InliningOptions.ShortMethod)]
         private void TransposeInplace_Scalar()
         {
-            float tmp;
-            int horIndex, verIndex;
-
-            // We don't care about the last row as it consists of a single element
-            // Which won't be swapped with anything
-            for (int i = 0; i < 7; i++)
+            ref float elemRef = ref Unsafe.As<Block8x8F, float>(ref this);
+
+            // row #0
+            Swap(ref Unsafe.Add(ref elemRef, 1), ref Unsafe.Add(ref elemRef, 8));
+            Swap(ref Unsafe.Add(ref elemRef, 2), ref Unsafe.Add(ref elemRef, 16));
+            Swap(ref Unsafe.Add(ref elemRef, 3), ref Unsafe.Add(ref elemRef, 24));
+            Swap(ref Unsafe.Add(ref elemRef, 4), ref Unsafe.Add(ref elemRef, 32));
+            Swap(ref Unsafe.Add(ref elemRef, 5), ref Unsafe.Add(ref elemRef, 40));
+            Swap(ref Unsafe.Add(ref elemRef, 6), ref Unsafe.Add(ref elemRef, 48));
+            Swap(ref Unsafe.Add(ref elemRef, 7), ref Unsafe.Add(ref elemRef, 56));
+
+            // row #1
+            Swap(ref Unsafe.Add(ref elemRef, 10), ref Unsafe.Add(ref elemRef, 17));
+            Swap(ref Unsafe.Add(ref elemRef, 11), ref Unsafe.Add(ref elemRef, 25));
+            Swap(ref Unsafe.Add(ref elemRef, 12), ref Unsafe.Add(ref elemRef, 33));
+            Swap(ref Unsafe.Add(ref elemRef, 13), ref Unsafe.Add(ref elemRef, 41));
+            Swap(ref Unsafe.Add(ref elemRef, 14), ref Unsafe.Add(ref elemRef, 49));
+            Swap(ref Unsafe.Add(ref elemRef, 15), ref Unsafe.Add(ref elemRef, 57));
+
+            // row #2
+            Swap(ref Unsafe.Add(ref elemRef, 19), ref Unsafe.Add(ref elemRef, 26));
+            Swap(ref Unsafe.Add(ref elemRef, 20), ref Unsafe.Add(ref elemRef, 34));
+            Swap(ref Unsafe.Add(ref elemRef, 21), ref Unsafe.Add(ref elemRef, 42));
+            Swap(ref Unsafe.Add(ref elemRef, 22), ref Unsafe.Add(ref elemRef, 50));
+            Swap(ref Unsafe.Add(ref elemRef, 23), ref Unsafe.Add(ref elemRef, 58));
+
+            // row #3
+            Swap(ref Unsafe.Add(ref elemRef, 28), ref Unsafe.Add(ref elemRef, 35));
+            Swap(ref Unsafe.Add(ref elemRef, 29), ref Unsafe.Add(ref elemRef, 43));
+            Swap(ref Unsafe.Add(ref elemRef, 30), ref Unsafe.Add(ref elemRef, 51));
+            Swap(ref Unsafe.Add(ref elemRef, 31), ref Unsafe.Add(ref elemRef, 59));
+
+            // row #4
+            Swap(ref Unsafe.Add(ref elemRef, 37), ref Unsafe.Add(ref elemRef, 44));
+            Swap(ref Unsafe.Add(ref elemRef, 38), ref Unsafe.Add(ref elemRef, 52));
+            Swap(ref Unsafe.Add(ref elemRef, 39), ref Unsafe.Add(ref elemRef, 60));
+
+            // row #5
+            Swap(ref Unsafe.Add(ref elemRef, 46), ref Unsafe.Add(ref elemRef, 53));
+            Swap(ref Unsafe.Add(ref elemRef, 47), ref Unsafe.Add(ref elemRef, 61));
+
+            // row #6
+            Swap(ref Unsafe.Add(ref elemRef, 55), ref Unsafe.Add(ref elemRef, 62));
+
+            static void Swap(ref float a, ref float b)
             {
-                // We don't care about the first element in each row as it's not swapped
-                for (int j = i + 1; j < 8; j++)
-                {
-                    horIndex = (i * 8) + j;
-                    verIndex = (j * 8) + i;
-                    tmp = this[horIndex];
-                    this[horIndex] = this[verIndex];
-                    this[verIndex] = tmp;
-                }
+                float tmp = a;
+                a = b;
+                b = tmp;
             }
         }
 

tests/ImageSharp.Benchmarks/Codecs/Jpeg/BlockOperations/Block8x8F_Transpose.cs

@@ -35,15 +35,18 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg.BlockOperations
 }
 
 /*
-BenchmarkDotNet=v0.13.0, OS=Windows 10.0.19042.1165 (20H2/October2020Update)
+BenchmarkDotNet=v0.13.0, OS=Windows 10.0.19042.1237 (20H2/October2020Update)
 Intel Core i7-6700K CPU 4.00GHz (Skylake), 1 CPU, 8 logical and 4 physical cores
 .NET SDK=6.0.100-preview.3.21202.5
-  [Host]          : .NET Core 3.1.18 (CoreCLR 4.700.21.35901, CoreFX 4.700.21.36305), X64 RyuJIT
-  AVX             : .NET Core 3.1.18 (CoreCLR 4.700.21.35901, CoreFX 4.700.21.36305), X64 RyuJIT
-  No HwIntrinsics : .NET Core 3.1.18 (CoreCLR 4.700.21.35901, CoreFX 4.700.21.36305), X64 RyuJIT
-
-|        Method |             Job |      Mean |     Error |    StdDev | Ratio | Gen 0 | Gen 1 | Gen 2 | Allocated |
-|-------------- |---------------- |----------:|----------:|----------:|------:|------:|------:|------:|----------:|
-| TransposeInto | No HwIntrinsics | 19.658 ns | 0.0550 ns | 0.0515 ns |  1.00 |     - |     - |     - |         - |
-| TransposeInto |             AVX |  8.613 ns | 0.0249 ns | 0.0208 ns |  0.44 |     - |     - |     - |         - |
+  [Host]             : .NET Core 3.1.18 (CoreCLR 4.700.21.35901, CoreFX 4.700.21.36305), X64 RyuJIT
+  1. No HwIntrinsics : .NET Core 3.1.18 (CoreCLR 4.700.21.35901, CoreFX 4.700.21.36305), X64 RyuJIT
+  2. SSE             : .NET Core 3.1.18 (CoreCLR 4.700.21.35901, CoreFX 4.700.21.36305), X64 RyuJIT
+  3. AVX             : .NET Core 3.1.18 (CoreCLR 4.700.21.35901, CoreFX 4.700.21.36305), X64 RyuJIT
+
+Runtime=.NET Core 3.1
+
+|           Method |             Job |      Mean |     Error |    StdDev | Ratio |
+|----------------- |----------------:|----------:|----------:|----------:|------:|
+| TransposeInplace | No HwIntrinsics | 12.531 ns | 0.0637 ns | 0.0565 ns |  1.00 |
+| TransposeInplace |             AVX |  5.767 ns | 0.0529 ns | 0.0495 ns |  0.46 |
 */