more cleanup + more tests

src/ImageSharp/Formats/Jpeg/Common/Block8x8F.Generated.cs

@@ -146,6 +146,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
                 
             ref Vector<float> row7 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V7L);
             row7 = NormalizeAndRound(row7, off, max);
-                        }
+                
+        }
 	}
 }

src/ImageSharp/Formats/Jpeg/Common/Block8x8F.Generated.tt

@@ -96,6 +96,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
                 <#
             }
 			#>
+
         }
 	}
 }

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

@@ -241,7 +241,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common
         /// </summary>
         /// <param name="dest">Destination</param>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public unsafe void CopyTo(Span<float> dest)
+        public void CopyTo(Span<float> dest)
         {
             ref byte d = ref Unsafe.As<float, byte>(ref dest.DangerousGetPinnableReference());
             ref byte s = ref Unsafe.As<Block8x8F, byte>(ref this);

src/ImageSharp/Formats/Jpeg/Common/Decoder/JpegBlockPostProcessor.cs

@@ -11,7 +11,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
     /// Encapsulates the implementation of processing "raw" <see cref="Buffer{T}"/>-s into Jpeg image channels.
     /// </summary>
     [StructLayout(LayoutKind.Sequential)]
-    internal unsafe struct JpegBlockPostProcessor
+    internal struct JpegBlockPostProcessor
     {
         /// <summary>
         /// Source block

src/ImageSharp/Formats/Jpeg/Common/Decoder/JpegColorConverter.FromYCbCr.cs

@@ -1,6 +1,7 @@
 using System;
 using System.Numerics;
 using System.Runtime.CompilerServices;
+using SixLabors.ImageSharp.Common.Tuples;
 
 namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
 {
@@ -124,9 +125,9 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
                     if (Vector<float>.Count == 4)
                     {
                         // TODO: Find a way to properly run & test this path on AVX2 PC-s! (Have I already mentioned that Vector<T> is terrible?)
-                        r.RoundAndDownscaleBasic();
+                        r.RoundAndDownscalePreAvx2();
-                        g.RoundAndDownscaleBasic();
+                        g.RoundAndDownscalePreAvx2();
-                        b.RoundAndDownscaleBasic();
+                        b.RoundAndDownscalePreAvx2();
                     }
                     else if (SimdUtils.IsAvx2CompatibleArchitecture)
                     {
@@ -147,67 +148,6 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Common.Decoder
                 }
             }
 
-            /// <summary>
-            /// Its faster to process multiple Vector4-s together
-            /// </summary>
-            private struct Vector4Pair
-            {
-                public Vector4 A;
-
-                public Vector4 B;
-
-                private static readonly Vector4 Scale = new Vector4(1 / 255f);
-
-                private static readonly Vector4 Half = new Vector4(0.5f);
-
-                [MethodImpl(MethodImplOptions.AggressiveInlining)]
-                public void RoundAndDownscaleBasic()
-                {
-                    ref Vector<float> a = ref Unsafe.As<Vector4, Vector<float>>(ref this.A);
-                    a = a.FastRound();
-
-                    ref Vector<float> b = ref Unsafe.As<Vector4, Vector<float>>(ref this.B);
-                    b = b.FastRound();
-
-                    // Downscale by 1/255
-                    this.A *= Scale;
-                    this.B *= Scale;
-                }
-
-                [MethodImpl(MethodImplOptions.AggressiveInlining)]
-                public void RoundAndDownscaleAvx2()
-                {
-                    ref Vector<float> self = ref Unsafe.As<Vector4Pair, Vector<float>>(ref this);
-                    Vector<float> v = self;
-                    v = v.FastRound();
-
-                    // Downscale by 1/255
-                    v *= new Vector<float>(1 / 255f);
-                    self = v;
-                }
-
-                [MethodImpl(MethodImplOptions.AggressiveInlining)]
-                public void MultiplyInplace(float value)
-                {
-                    this.A *= value;
-                    this.B *= value;
-                }
-
-                [MethodImpl(MethodImplOptions.AggressiveInlining)]
-                public void AddInplace(Vector4 value)
-                {
-                    this.A += value;
-                    this.B += value;
-                }
-
-                [MethodImpl(MethodImplOptions.AggressiveInlining)]
-                public void AddInplace(ref Vector4Pair other)
-                {
-                    this.A += other.A;
-                    this.B += other.B;
-                }
-            }
-
             private struct Vector4Octet
             {
 #pragma warning disable SA1132 // Do not combine fields

tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs

@@ -32,6 +32,16 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
         {
         }
 
+        private bool SkipOnNonAvx2Runner()
+        {
+            if (!SimdUtils.IsAvx2CompatibleArchitecture)
+            {
+                this.Output.WriteLine("AVX2 not supported, skipping!");
+                return true;
+            }
+            return false;
+        }
+
         [Fact]
         public void Indexer()
         {
@@ -263,9 +273,8 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
         [InlineData(2)]
         public void NormalizeColorsAndRoundAvx2(int seed)
         {
-            if (!SimdUtils.IsAvx2CompatibleArchitecture)
+            if (this.SkipOnNonAvx2Runner())
             {
-                this.Output.WriteLine("AVX2 not supported, skipping!");
                 return;
             }
 
@@ -283,6 +292,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             this.CompareBlocks(expected, actual, 0);
         }
 
+
         [Theory]
         [InlineData(1)]
         [InlineData(2)]
@@ -413,5 +423,19 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
             this.CompareBlocks(expected, actual, 0);
         }
+
+        [Fact]
+        public void MultiplyInplace_ByScalar()
+        {
+            Block8x8F original = CreateRandomFloatBlock(-500, 500);
+
+            Block8x8F actual = original;
+            actual.MultiplyInplace(42f);
+
+            for (int i = 0; i < 64; i++)
+            {
+                Assert.Equal(original[i]*42f, actual[i]);
+            }
+        }
     }
 }