Merge pull request #1143 from SixLabors/af/block-scale-optimization

Undo jpeg perf regression, add various optimizations

src/ImageSharp/Common/Helpers/SimdUtils.Avx2Intrinsics.cs

@@ -0,0 +1,103 @@
+// Copyright (c) Six Labors and contributors.
+// Licensed under the Apache License, Version 2.0.
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+
+using System;
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+
+namespace SixLabors.ImageSharp
+{
+    internal static partial class SimdUtils
+    {
+        public static class Avx2Intrinsics
+        {
+            private static ReadOnlySpan<byte> PermuteMaskDeinterleave8x32 => new byte[] { 0, 0, 0, 0, 4, 0, 0, 0, 1, 0, 0, 0, 5, 0, 0, 0, 2, 0, 0, 0, 6, 0, 0, 0, 3, 0, 0, 0, 7, 0, 0, 0 };
+
+            /// <summary>
+            /// <see cref="NormalizedFloatToByteSaturate"/> as many elements as possible, slicing them down (keeping the remainder).
+            /// </summary>
+            [MethodImpl(InliningOptions.ShortMethod)]
+            internal static void NormalizedFloatToByteSaturateReduce(
+                ref ReadOnlySpan<float> source,
+                ref Span<byte> dest)
+            {
+                DebugGuard.IsTrue(source.Length == dest.Length, nameof(source), "Input spans must be of same length!");
+
+                if (Avx2.IsSupported)
+                {
+                    int remainder = ImageMaths.ModuloP2(source.Length, Vector<byte>.Count);
+                    int adjustedCount = source.Length - remainder;
+
+                    if (adjustedCount > 0)
+                    {
+                        NormalizedFloatToByteSaturate(
+                            source.Slice(0, adjustedCount),
+                            dest.Slice(0, adjustedCount));
+
+                        source = source.Slice(adjustedCount);
+                        dest = dest.Slice(adjustedCount);
+                    }
+                }
+            }
+
+            /// <summary>
+            /// Implementation of <see cref="SimdUtils.NormalizedFloatToByteSaturate"/>, which is faster on new .NET runtime.
+            /// </summary>
+            /// <remarks>
+            /// Implementation is based on MagicScaler code:
+            /// https://github.com/saucecontrol/PhotoSauce/blob/a9bd6e5162d2160419f0cf743fd4f536c079170b/src/MagicScaler/Magic/Processors/ConvertersFloat.cs#L453-L477
+            /// </remarks>
+            internal static void NormalizedFloatToByteSaturate(
+                ReadOnlySpan<float> source,
+                Span<byte> dest)
+            {
+                VerifySpanInput(source, dest, Vector256<byte>.Count);
+
+                int n = dest.Length / Vector256<byte>.Count;
+
+                ref Vector256<float> sourceBase =
+                    ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(source));
+                ref Vector256<byte> destBase = ref Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(dest));
+
+                var maxBytes = Vector256.Create(255f);
+                ref byte maskBase = ref MemoryMarshal.GetReference(PermuteMaskDeinterleave8x32);
+                Vector256<int> mask = Unsafe.As<byte, Vector256<int>>(ref maskBase);
+
+                for (int i = 0; i < n; i++)
+                {
+                    ref Vector256<float> s = ref Unsafe.Add(ref sourceBase, i * 4);
+
+                    Vector256<float> f0 = s;
+                    Vector256<float> f1 = Unsafe.Add(ref s, 1);
+                    Vector256<float> f2 = Unsafe.Add(ref s, 2);
+                    Vector256<float> f3 = Unsafe.Add(ref s, 3);
+
+                    Vector256<int> w0 = ConvertToInt32(f0, maxBytes);
+                    Vector256<int> w1 = ConvertToInt32(f1, maxBytes);
+                    Vector256<int> w2 = ConvertToInt32(f2, maxBytes);
+                    Vector256<int> w3 = ConvertToInt32(f3, maxBytes);
+
+                    Vector256<short> u0 = Avx2.PackSignedSaturate(w0, w1);
+                    Vector256<short> u1 = Avx2.PackSignedSaturate(w2, w3);
+                    Vector256<byte> b = Avx2.PackUnsignedSaturate(u0, u1);
+                    b = Avx2.PermuteVar8x32(b.AsInt32(), mask).AsByte();
+
+                    Unsafe.Add(ref destBase, i) = b;
+                }
+            }
+
+            [MethodImpl(MethodImplOptions.AggressiveInlining)]
+            private static Vector256<int> ConvertToInt32(Vector256<float> vf, Vector256<float> scale)
+            {
+                vf = Avx.Multiply(vf, scale);
+                return Avx.ConvertToVector256Int32(vf);
+            }
+        }
+    }
+}
+#endif

src/ImageSharp/Common/Helpers/SimdUtils.BasicIntrinsics256.cs

@@ -17,14 +17,14 @@ namespace SixLabors.ImageSharp
         /// </summary>
         public static class BasicIntrinsics256
         {
-            public static bool IsAvailable { get; } = IsAvx2CompatibleArchitecture;
+            public static bool IsAvailable { get; } = HasVector8;
 
 #if !SUPPORTS_EXTENDED_INTRINSICS
             /// <summary>
-            /// <see cref="BulkConvertByteToNormalizedFloat"/> as many elements as possible, slicing them down (keeping the remainder).
+            /// <see cref="ByteToNormalizedFloat"/> as many elements as possible, slicing them down (keeping the remainder).
             /// </summary>
             [MethodImpl(InliningOptions.ShortMethod)]
-            internal static void BulkConvertByteToNormalizedFloatReduce(
+            internal static void ByteToNormalizedFloatReduce(
                 ref ReadOnlySpan<byte> source,
                 ref Span<float> dest)
             {
@@ -40,7 +40,7 @@ namespace SixLabors.ImageSharp
 
                 if (adjustedCount > 0)
                 {
-                    BulkConvertByteToNormalizedFloat(
+                    ByteToNormalizedFloat(
                         source.Slice(0, adjustedCount),
                         dest.Slice(0, adjustedCount));
 
@@ -50,10 +50,10 @@ namespace SixLabors.ImageSharp
             }
 
             /// <summary>
-            /// <see cref="BulkConvertNormalizedFloatToByteClampOverflows"/> as many elements as possible, slicing them down (keeping the remainder).
+            /// <see cref="NormalizedFloatToByteSaturate"/> as many elements as possible, slicing them down (keeping the remainder).
             /// </summary>
             [MethodImpl(InliningOptions.ShortMethod)]
-            internal static void BulkConvertNormalizedFloatToByteClampOverflowsReduce(
+            internal static void NormalizedFloatToByteSaturateReduce(
                 ref ReadOnlySpan<float> source,
                 ref Span<byte> dest)
             {
@@ -69,7 +69,7 @@ namespace SixLabors.ImageSharp
 
                 if (adjustedCount > 0)
                 {
-                    BulkConvertNormalizedFloatToByteClampOverflows(source.Slice(0, adjustedCount), dest.Slice(0, adjustedCount));
+                    NormalizedFloatToByteSaturate(source.Slice(0, adjustedCount), dest.Slice(0, adjustedCount));
 
                     source = source.Slice(adjustedCount);
                     dest = dest.Slice(adjustedCount);
@@ -78,15 +78,15 @@ namespace SixLabors.ImageSharp
 #endif
 
             /// <summary>
-            /// SIMD optimized implementation for <see cref="SimdUtils.BulkConvertByteToNormalizedFloat"/>.
+            /// SIMD optimized implementation for <see cref="SimdUtils.ByteToNormalizedFloat"/>.
             /// Works only with span Length divisible by 8.
             /// Implementation adapted from:
             /// http://lolengine.net/blog/2011/3/20/understanding-fast-float-integer-conversions
             /// http://stackoverflow.com/a/536278
             /// </summary>
-            internal static void BulkConvertByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
+            internal static void ByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
             {
-                VerifyIsAvx2Compatible(nameof(BulkConvertByteToNormalizedFloat));
+                VerifyHasVector8(nameof(ByteToNormalizedFloat));
                 VerifySpanInput(source, dest, 8);
 
                 var bVec = new Vector<float>(256.0f / 255.0f);
@@ -124,11 +124,11 @@ namespace SixLabors.ImageSharp
             }
 
             /// <summary>
-            /// Implementation of <see cref="SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows"/> which is faster on older runtimes.
+            /// Implementation of <see cref="SimdUtils.NormalizedFloatToByteSaturate"/> which is faster on older runtimes.
             /// </summary>
-            internal static void BulkConvertNormalizedFloatToByteClampOverflows(ReadOnlySpan<float> source, Span<byte> dest)
+            internal static void NormalizedFloatToByteSaturate(ReadOnlySpan<float> source, Span<byte> dest)
             {
-                VerifyIsAvx2Compatible(nameof(BulkConvertNormalizedFloatToByteClampOverflows));
+                VerifyHasVector8(nameof(NormalizedFloatToByteSaturate));
                 VerifySpanInput(source, dest, 8);
 
                 if (source.Length == 0)
@@ -177,7 +177,7 @@ namespace SixLabors.ImageSharp
             /// </summary>
             internal static void BulkConvertNormalizedFloatToByte(ReadOnlySpan<float> source, Span<byte> dest)
             {
-                VerifyIsAvx2Compatible(nameof(BulkConvertNormalizedFloatToByte));
+                VerifyHasVector8(nameof(BulkConvertNormalizedFloatToByte));
                 VerifySpanInput(source, dest, 8);
 
                 if (source.Length == 0)

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

@@ -43,10 +43,10 @@ namespace SixLabors.ImageSharp
             }
 
             /// <summary>
-            /// <see cref="BulkConvertByteToNormalizedFloat"/> as many elements as possible, slicing them down (keeping the remainder).
+            /// <see cref="ByteToNormalizedFloat"/> as many elements as possible, slicing them down (keeping the remainder).
             /// </summary>
             [MethodImpl(InliningOptions.ShortMethod)]
-            internal static void BulkConvertByteToNormalizedFloatReduce(
+            internal static void ByteToNormalizedFloatReduce(
                 ref ReadOnlySpan<byte> source,
                 ref Span<float> dest)
             {
@@ -62,7 +62,7 @@ namespace SixLabors.ImageSharp
 
                 if (adjustedCount > 0)
                 {
-                    BulkConvertByteToNormalizedFloat(source.Slice(0, adjustedCount), dest.Slice(0, adjustedCount));
+                    ByteToNormalizedFloat(source.Slice(0, adjustedCount), dest.Slice(0, adjustedCount));
 
                     source = source.Slice(adjustedCount);
                     dest = dest.Slice(adjustedCount);
@@ -70,10 +70,10 @@ namespace SixLabors.ImageSharp
             }
 
             /// <summary>
-            /// <see cref="BulkConvertNormalizedFloatToByteClampOverflows"/> as many elements as possible, slicing them down (keeping the remainder).
+            /// <see cref="NormalizedFloatToByteSaturate"/> as many elements as possible, slicing them down (keeping the remainder).
             /// </summary>
             [MethodImpl(InliningOptions.ShortMethod)]
-            internal static void BulkConvertNormalizedFloatToByteClampOverflowsReduce(
+            internal static void NormalizedFloatToByteSaturateReduce(
                 ref ReadOnlySpan<float> source,
                 ref Span<byte> dest)
             {
@@ -89,7 +89,7 @@ namespace SixLabors.ImageSharp
 
                 if (adjustedCount > 0)
                 {
-                    BulkConvertNormalizedFloatToByteClampOverflows(
+                    NormalizedFloatToByteSaturate(
                         source.Slice(0, adjustedCount),
                         dest.Slice(0, adjustedCount));
 
@@ -99,9 +99,9 @@ namespace SixLabors.ImageSharp
             }
 
             /// <summary>
-            /// Implementation <see cref="SimdUtils.BulkConvertByteToNormalizedFloat"/>, which is faster on new RyuJIT runtime.
+            /// Implementation <see cref="SimdUtils.ByteToNormalizedFloat"/>, which is faster on new RyuJIT runtime.
             /// </summary>
-            internal static void BulkConvertByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
+            internal static void ByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
             {
                 VerifySpanInput(source, dest, Vector<byte>.Count);
 
@@ -132,9 +132,9 @@ namespace SixLabors.ImageSharp
             }
 
             /// <summary>
-            /// Implementation of <see cref="SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows"/>, which is faster on new .NET runtime.
+            /// Implementation of <see cref="SimdUtils.NormalizedFloatToByteSaturate"/>, which is faster on new .NET runtime.
             /// </summary>
-            internal static void BulkConvertNormalizedFloatToByteClampOverflows(
+            internal static void NormalizedFloatToByteSaturate(
                 ReadOnlySpan<float> source,
                 Span<byte> dest)
             {

src/ImageSharp/Common/Helpers/SimdUtils.FallbackIntrinsics128.cs

@@ -19,10 +19,10 @@ namespace SixLabors.ImageSharp
         public static class FallbackIntrinsics128
         {
             /// <summary>
-            /// <see cref="BulkConvertByteToNormalizedFloat"/> as many elements as possible, slicing them down (keeping the remainder).
+            /// <see cref="ByteToNormalizedFloat"/> as many elements as possible, slicing them down (keeping the remainder).
             /// </summary>
             [MethodImpl(InliningOptions.ShortMethod)]
-            internal static void BulkConvertByteToNormalizedFloatReduce(
+            internal static void ByteToNormalizedFloatReduce(
                 ref ReadOnlySpan<byte> source,
                 ref Span<float> dest)
             {
@@ -33,7 +33,7 @@ namespace SixLabors.ImageSharp
 
                 if (adjustedCount > 0)
                 {
-                    BulkConvertByteToNormalizedFloat(
+                    ByteToNormalizedFloat(
                         source.Slice(0, adjustedCount),
                         dest.Slice(0, adjustedCount));
 
@@ -43,10 +43,10 @@ namespace SixLabors.ImageSharp
             }
 
             /// <summary>
-            /// <see cref="BulkConvertNormalizedFloatToByteClampOverflows"/> as many elements as possible, slicing them down (keeping the remainder).
+            /// <see cref="NormalizedFloatToByteSaturate"/> as many elements as possible, slicing them down (keeping the remainder).
             /// </summary>
             [MethodImpl(InliningOptions.ShortMethod)]
-            internal static void BulkConvertNormalizedFloatToByteClampOverflowsReduce(
+            internal static void NormalizedFloatToByteSaturateReduce(
                 ref ReadOnlySpan<float> source,
                 ref Span<byte> dest)
             {
@@ -57,7 +57,7 @@ namespace SixLabors.ImageSharp
 
                 if (adjustedCount > 0)
                 {
-                    BulkConvertNormalizedFloatToByteClampOverflows(
+                    NormalizedFloatToByteSaturate(
                         source.Slice(0, adjustedCount),
                         dest.Slice(0, adjustedCount));
 
@@ -67,10 +67,10 @@ namespace SixLabors.ImageSharp
             }
 
             /// <summary>
-            /// Implementation of <see cref="SimdUtils.BulkConvertByteToNormalizedFloat"/> using <see cref="Vector4"/>.
+            /// Implementation of <see cref="SimdUtils.ByteToNormalizedFloat"/> using <see cref="Vector4"/>.
             /// </summary>
             [MethodImpl(InliningOptions.ColdPath)]
-            internal static void BulkConvertByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
+            internal static void ByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
             {
                 VerifySpanInput(source, dest, 4);
 
@@ -99,10 +99,10 @@ namespace SixLabors.ImageSharp
             }
 
             /// <summary>
-            /// Implementation of <see cref="SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows"/> using <see cref="Vector4"/>.
+            /// Implementation of <see cref="SimdUtils.NormalizedFloatToByteSaturate"/> using <see cref="Vector4"/>.
             /// </summary>
             [MethodImpl(InliningOptions.ColdPath)]
-            internal static void BulkConvertNormalizedFloatToByteClampOverflows(
+            internal static void NormalizedFloatToByteSaturate(
                 ReadOnlySpan<float> source,
                 Span<byte> dest)
             {
@@ -148,4 +148,4 @@ namespace SixLabors.ImageSharp
             }
         }
     }
-}
+}

src/ImageSharp/Common/Helpers/SimdUtils.cs

@@ -15,9 +15,10 @@ namespace SixLabors.ImageSharp
     internal static partial class SimdUtils
     {
         /// <summary>
-        /// Gets a value indicating whether the code is being executed on AVX2 CPU where both float and integer registers are of size 256 byte.
+        /// Gets a value indicating whether <see cref="Vector{T}"/> code is being JIT-ed to AVX2 instructions
+        /// where both float and integer registers are of size 256 byte.
         /// </summary>
-        public static bool IsAvx2CompatibleArchitecture { get; } =
+        public static bool HasVector8 { get; } =
             Vector.IsHardwareAccelerated && Vector<float>.Count == 8 && Vector<int>.Count == 8;
 
         /// <summary>
@@ -60,16 +61,18 @@ namespace SixLabors.ImageSharp
         /// <param name="source">The source span of bytes</param>
         /// <param name="dest">The destination span of floats</param>
         [MethodImpl(InliningOptions.ShortMethod)]
-        internal static void BulkConvertByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
+        internal static void ByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
         {
             DebugGuard.IsTrue(source.Length == dest.Length, nameof(source), "Input spans must be of same length!");
 
 #if SUPPORTS_EXTENDED_INTRINSICS
-            ExtendedIntrinsics.BulkConvertByteToNormalizedFloatReduce(ref source, ref dest);
+            ExtendedIntrinsics.ByteToNormalizedFloatReduce(ref source, ref dest);
 #else
-            BasicIntrinsics256.BulkConvertByteToNormalizedFloatReduce(ref source, ref dest);
+            BasicIntrinsics256.ByteToNormalizedFloatReduce(ref source, ref dest);
 #endif
-            FallbackIntrinsics128.BulkConvertByteToNormalizedFloatReduce(ref source, ref dest);
+
+            // Also deals with the remainder from previous conversions:
+            FallbackIntrinsics128.ByteToNormalizedFloatReduce(ref source, ref dest);
 
             // Deal with the remainder:
             if (source.Length > 0)
@@ -87,16 +90,20 @@ namespace SixLabors.ImageSharp
         /// <param name="source">The source span of floats</param>
         /// <param name="dest">The destination span of bytes</param>
         [MethodImpl(InliningOptions.ShortMethod)]
-        internal static void BulkConvertNormalizedFloatToByteClampOverflows(ReadOnlySpan<float> source, Span<byte> dest)
+        internal static void NormalizedFloatToByteSaturate(ReadOnlySpan<float> source, Span<byte> dest)
         {
             DebugGuard.IsTrue(source.Length == dest.Length, nameof(source), "Input spans must be of same length!");
 
-#if SUPPORTS_EXTENDED_INTRINSICS
-            ExtendedIntrinsics.BulkConvertNormalizedFloatToByteClampOverflowsReduce(ref source, ref dest);
+#if SUPPORTS_RUNTIME_INTRINSICS
+            Avx2Intrinsics.NormalizedFloatToByteSaturateReduce(ref source, ref dest);
+#elif SUPPORTS_EXTENDED_INTRINSICS
+            ExtendedIntrinsics.NormalizedFloatToByteSaturateReduce(ref source, ref dest);
 #else
-            BasicIntrinsics256.BulkConvertNormalizedFloatToByteClampOverflowsReduce(ref source, ref dest);
+            BasicIntrinsics256.NormalizedFloatToByteSaturateReduce(ref source, ref dest);
 #endif
-            FallbackIntrinsics128.BulkConvertNormalizedFloatToByteClampOverflowsReduce(ref source, ref dest);
+
+            // Also deals with the remainder from previous conversions:
+            FallbackIntrinsics128.NormalizedFloatToByteSaturateReduce(ref source, ref dest);
 
             // Deal with the remainder:
             if (source.Length > 0)
@@ -151,9 +158,9 @@ namespace SixLabors.ImageSharp
         private static byte ConvertToByte(float f) => (byte)ComparableExtensions.Clamp((f * 255f) + 0.5f, 0, 255f);
 
         [Conditional("DEBUG")]
-        private static void VerifyIsAvx2Compatible(string operation)
+        private static void VerifyHasVector8(string operation)
         {
-            if (!IsAvx2CompatibleArchitecture)
+            if (!HasVector8)
             {
                 throw new NotSupportedException($"{operation} is supported only on AVX2 CPU!");
             }

src/ImageSharp/Common/Tuples/Vector4Pair.cs

@@ -44,7 +44,7 @@ namespace SixLabors.ImageSharp.Tuples
         /// Downscale method, specific to Jpeg color conversion. Works only if Vector{float}.Count == 4!        /// TODO: Move it somewhere else.
         /// </summary>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        internal void RoundAndDownscalePreAvx2(float downscaleFactor)
+        internal void RoundAndDownscalePreVector8(float downscaleFactor)
         {
             ref Vector<float> a = ref Unsafe.As<Vector4, Vector<float>>(ref this.A);
             a = a.FastRound();
@@ -63,7 +63,7 @@ namespace SixLabors.ImageSharp.Tuples
         /// TODO: Move it somewhere else.
         /// </summary>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        internal void RoundAndDownscaleAvx2(float downscaleFactor)
+        internal void RoundAndDownscaleVector8(float downscaleFactor)
         {
             ref Vector<float> self = ref Unsafe.As<Vector4Pair, Vector<float>>(ref this);
             Vector<float> v = self;
@@ -79,4 +79,4 @@ namespace SixLabors.ImageSharp.Tuples
             return $"{nameof(Vector4Pair)}({this.A}, {this.B})";
         }
     }
-}
+}

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

@@ -121,7 +121,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         /// AVX2-only variant for executing <see cref="NormalizeColorsInplace"/> and <see cref="RoundInplace"/> in one step.
         /// </summary>
         [MethodImpl(InliningOptions.ShortMethod)]
-        public void NormalizeColorsAndRoundInplaceAvx2(float maximum)
+        public void NormalizeColorsAndRoundInplaceVector8(float maximum)
         {
             var off = new Vector<float>(MathF.Ceiling(maximum / 2));
             var max = new Vector<float>(maximum);

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

@@ -84,7 +84,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         /// AVX2-only variant for executing <see cref="NormalizeColorsInplace"/> and <see cref="RoundInplace"/> in one step.
         /// </summary>
         [MethodImpl(InliningOptions.ShortMethod)]
-        public void NormalizeColorsAndRoundInplaceAvx2(float maximum)
+        public void NormalizeColorsAndRoundInplaceVector8(float maximum)
         {
             var off = new Vector<float>(MathF.Ceiling(maximum / 2));
             var max = new Vector<float>(maximum);

src/ImageSharp/Formats/Jpeg/Components/Block8x8F.CopyTo.cs → src/ImageSharp/Formats/Jpeg/Components/Block8x8F.ScaledCopyTo.cs

@@ -15,29 +15,36 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         /// Copy block data into the destination color buffer pixel area with the provided horizontal and vertical scale factors.
         /// </summary>
         [MethodImpl(InliningOptions.ShortMethod)]
-        public void CopyTo(in BufferArea<float> area, int horizontalScale, int verticalScale)
+        public void ScaledCopyTo(in BufferArea<float> area, int horizontalScale, int verticalScale)
+        {
+            ref float areaOrigin = ref area.GetReferenceToOrigin();
+            this.ScaledCopyTo(ref areaOrigin, area.Stride, horizontalScale, verticalScale);
+        }
+
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public void ScaledCopyTo(ref float areaOrigin, int areaStride, int horizontalScale, int verticalScale)
         {
             if (horizontalScale == 1 && verticalScale == 1)
             {
-                this.Copy1x1Scale(area);
+                this.Copy1x1Scale(ref areaOrigin, areaStride);
                 return;
             }
 
             if (horizontalScale == 2 && verticalScale == 2)
             {
-                this.Copy2x2Scale(area);
+                this.Copy2x2Scale(ref areaOrigin, areaStride);
                 return;
             }
 
             // TODO: Optimize: implement all cases with scale-specific, loopless code!
-            this.CopyArbitraryScale(area, horizontalScale, verticalScale);
+            this.CopyArbitraryScale(ref areaOrigin, areaStride, horizontalScale, verticalScale);
         }
 
-        public void Copy1x1Scale(in BufferArea<float> destination)
+        public void Copy1x1Scale(ref float areaOrigin, int areaStride)
         {
             ref byte selfBase = ref Unsafe.As<Block8x8F, byte>(ref this);
-            ref byte destBase = ref Unsafe.As<float, byte>(ref destination.GetReferenceToOrigin());
-            int destStride = destination.Stride * sizeof(float);
+            ref byte destBase = ref Unsafe.As<float, byte>(ref areaOrigin);
+            int destStride = areaStride * sizeof(float);
 
             CopyRowImpl(ref selfBase, ref destBase, destStride, 0);
             CopyRowImpl(ref selfBase, ref destBase, destStride, 1);
@@ -57,10 +64,10 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
             Unsafe.CopyBlock(ref d, ref s, 8 * sizeof(float));
         }
 
-        private void Copy2x2Scale(in BufferArea<float> area)
+        private void Copy2x2Scale(ref float areaOrigin, int areaStride)
         {
-            ref Vector2 destBase = ref Unsafe.As<float, Vector2>(ref area.GetReferenceToOrigin());
-            int destStride = area.Stride / 2;
+            ref Vector2 destBase = ref Unsafe.As<float, Vector2>(ref areaOrigin);
+            int destStride = areaStride / 2;
 
             this.WidenCopyRowImpl2x2(ref destBase, 0, destStride);
             this.WidenCopyRowImpl2x2(ref destBase, 1, destStride);
@@ -110,10 +117,8 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         }
 
         [MethodImpl(InliningOptions.ColdPath)]
-        private void CopyArbitraryScale(BufferArea<float> area, int horizontalScale, int verticalScale)
+        private void CopyArbitraryScale(ref float areaOrigin, int areaStride, int horizontalScale, int verticalScale)
         {
-            ref float destBase = ref area.GetReferenceToOrigin();
-
             for (int y = 0; y < 8; y++)
             {
                 int yy = y * verticalScale;
@@ -127,12 +132,12 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
 
                     for (int i = 0; i < verticalScale; i++)
                     {
-                        int baseIdx = ((yy + i) * area.Stride) + xx;
+                        int baseIdx = ((yy + i) * areaStride) + xx;
 
                         for (int j = 0; j < horizontalScale; j++)
                         {
                             // area[xx + j, yy + i] = value;
-                            Unsafe.Add(ref destBase, baseIdx + j) = value;
+                            Unsafe.Add(ref areaOrigin, baseIdx + j) = value;
                         }
                     }
                 }

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

@@ -201,7 +201,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         /// </summary>
         /// <param name="dest">Destination</param>
         [MethodImpl(InliningOptions.ShortMethod)]
-        public void CopyTo(Span<float> dest)
+        public void ScaledCopyTo(Span<float> dest)
         {
             ref byte d = ref Unsafe.As<float, byte>(ref MemoryMarshal.GetReference(dest));
             ref byte s = ref Unsafe.As<Block8x8F, byte>(ref this);
@@ -215,7 +215,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         /// <param name="blockPtr">Pointer to block</param>
         /// <param name="dest">Destination</param>
         [MethodImpl(InliningOptions.ShortMethod)]
-        public static unsafe void CopyTo(Block8x8F* blockPtr, Span<byte> dest)
+        public static unsafe void ScaledCopyTo(Block8x8F* blockPtr, Span<byte> dest)
         {
             float* fPtr = (float*)blockPtr;
             for (int i = 0; i < Size; i++)
@@ -231,9 +231,9 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         /// <param name="blockPtr">The block pointer.</param>
         /// <param name="dest">The destination.</param>
         [MethodImpl(InliningOptions.ShortMethod)]
-        public static unsafe void CopyTo(Block8x8F* blockPtr, Span<float> dest)
+        public static unsafe void ScaledCopyTo(Block8x8F* blockPtr, Span<float> dest)
         {
-            blockPtr->CopyTo(dest);
+            blockPtr->ScaledCopyTo(dest);
         }
 
         /// <summary>
@@ -241,7 +241,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         /// </summary>
         /// <param name="dest">Destination</param>
         [MethodImpl(InliningOptions.ShortMethod)]
-        public unsafe void CopyTo(float[] dest)
+        public unsafe void ScaledCopyTo(float[] dest)
         {
             fixed (void* ptr = &this.V0L)
             {
@@ -253,7 +253,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         /// Copy raw 32bit floating point data to dest
         /// </summary>
         /// <param name="dest">Destination</param>
-        public unsafe void CopyTo(Span<int> dest)
+        public unsafe void ScaledCopyTo(Span<int> dest)
         {
             fixed (Vector4* ptr = &this.V0L)
             {
@@ -268,7 +268,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         public float[] ToArray()
         {
             var result = new float[Size];
-            this.CopyTo(result);
+            this.ScaledCopyTo(result);
             return result;
         }
 
@@ -471,9 +471,9 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         /// </summary>
         public void NormalizeColorsAndRoundInplace(float maximum)
         {
-            if (SimdUtils.IsAvx2CompatibleArchitecture)
+            if (SimdUtils.HasVector8)
             {
-                this.NormalizeColorsAndRoundInplaceAvx2(maximum);
+                this.NormalizeColorsAndRoundInplaceVector8(maximum);
             }
             else
             {
@@ -497,7 +497,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         public void LoadFrom(ref Block8x8 source)
         {
 #if SUPPORTS_EXTENDED_INTRINSICS
-            if (SimdUtils.IsAvx2CompatibleArchitecture)
+            if (SimdUtils.HasVector8)
             {
                 this.LoadFromInt16ExtendedAvx2(ref source);
                 return;
@@ -513,7 +513,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components
         public void LoadFromInt16ExtendedAvx2(ref Block8x8 source)
         {
             DebugGuard.IsTrue(
-                SimdUtils.IsAvx2CompatibleArchitecture,
+                SimdUtils.HasVector8,
                 "LoadFromUInt16ExtendedAvx2 only works on AVX2 compatible architecture!");
 
             ref Vector<short> sRef = ref Unsafe.As<Block8x8, Vector<short>>(ref source);

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromYCbCrSimd.cs

@@ -90,15 +90,15 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
                     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.RoundAndDownscalePreAvx2(maxValue);
-                        g.RoundAndDownscalePreAvx2(maxValue);
-                        b.RoundAndDownscalePreAvx2(maxValue);
+                        r.RoundAndDownscalePreVector8(maxValue);
+                        g.RoundAndDownscalePreVector8(maxValue);
+                        b.RoundAndDownscalePreVector8(maxValue);
                     }
-                    else if (SimdUtils.IsAvx2CompatibleArchitecture)
+                    else if (SimdUtils.HasVector8)
                     {
-                        r.RoundAndDownscaleAvx2(maxValue);
-                        g.RoundAndDownscaleAvx2(maxValue);
-                        b.RoundAndDownscaleAvx2(maxValue);
+                        r.RoundAndDownscaleVector8(maxValue);
+                        g.RoundAndDownscaleVector8(maxValue);
+                        b.RoundAndDownscaleVector8(maxValue);
                     }
                     else
                     {
@@ -114,4 +114,4 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
             }
         }
     }
-}
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.FromYCbCrSimdAvx2.cs

@@ -13,14 +13,14 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
 {
     internal abstract partial class JpegColorConverter
     {
-        internal sealed class FromYCbCrSimdAvx2 : JpegColorConverter
+        internal sealed class FromYCbCrSimdVector8 : JpegColorConverter
         {
-            public FromYCbCrSimdAvx2(int precision)
+            public FromYCbCrSimdVector8(int precision)
                 : base(JpegColorSpace.YCbCr, precision)
             {
             }
 
-            public static bool IsAvailable => Vector.IsHardwareAccelerated && SimdUtils.IsAvx2CompatibleArchitecture;
+            public static bool IsAvailable => Vector.IsHardwareAccelerated && SimdUtils.HasVector8;
 
             public override void ConvertToRgba(in ComponentValues values, Span<Vector4> result)
             {
@@ -107,4 +107,4 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
             }
         }
     }
-}
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/ColorConverters/JpegColorConverter.cs

@@ -93,7 +93,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters
         /// Returns the <see cref="JpegColorConverter"/> for the YCbCr colorspace that matches the current CPU architecture.
         /// </summary>
         private static JpegColorConverter GetYCbCrConverter(int precision) =>
-            FromYCbCrSimdAvx2.IsAvailable ? (JpegColorConverter)new FromYCbCrSimdAvx2(precision) : new FromYCbCrSimd(precision);
+            FromYCbCrSimdVector8.IsAvailable ? (JpegColorConverter)new FromYCbCrSimdVector8(precision) : new FromYCbCrSimd(precision);
 
         /// <summary>
         /// A stack-only struct to reference the input buffers using <see cref="ReadOnlySpan{T}"/>-s.

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

@@ -68,11 +68,13 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
         /// - Copy the resulting color values into 'destArea' scaling up the block by amount defined in <see cref="subSamplingDivisors"/>.
         /// </summary>
         /// <param name="sourceBlock">The source block.</param>
-        /// <param name="destArea">The destination buffer area.</param>
+        /// <param name="destAreaOrigin">Reference to the origin of the destination pixel area.</param>
+        /// <param name="destAreaStride">The width of the destination pixel buffer.</param>
         /// <param name="maximumValue">The maximum value derived from the bitdepth.</param>
         public void ProcessBlockColorsInto(
             ref Block8x8 sourceBlock,
-            in BufferArea<float> destArea,
+            ref float destAreaOrigin,
+            int destAreaStride,
             float maximumValue)
         {
             ref Block8x8F b = ref this.SourceBlock;
@@ -88,7 +90,11 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
             // To be "more accurate", we need to emulate this by rounding!
             this.WorkspaceBlock1.NormalizeColorsAndRoundInplace(maximumValue);
 
-            this.WorkspaceBlock1.CopyTo(destArea, this.subSamplingDivisors.Width, this.subSamplingDivisors.Height);
+            this.WorkspaceBlock1.ScaledCopyTo(
+                ref destAreaOrigin,
+                destAreaStride,
+                this.subSamplingDivisors.Width,
+                this.subSamplingDivisors.Height);
         }
     }
-}
+}

src/ImageSharp/Formats/Jpeg/Components/Decoder/JpegComponentPostProcessor.cs

@@ -79,6 +79,8 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
             var blockPp = new JpegBlockPostProcessor(this.ImagePostProcessor.RawJpeg, this.Component);
             float maximumValue = MathF.Pow(2, this.ImagePostProcessor.RawJpeg.Precision) - 1;
 
+            int destAreaStride = this.ColorBuffer.Width;
+
             for (int y = 0; y < this.BlockRowsPerStep; y++)
             {
                 int yBlock = this.currentComponentRowInBlocks + y;
@@ -90,22 +92,19 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder
 
                 int yBuffer = y * this.blockAreaSize.Height;
 
+                Span<float> colorBufferRow = this.ColorBuffer.GetRowSpan(yBuffer);
                 Span<Block8x8> blockRow = this.Component.SpectralBlocks.GetRowSpan(yBlock);
 
-                ref Block8x8 blockRowBase = ref MemoryMarshal.GetReference(blockRow);
+                // see: https://github.com/SixLabors/ImageSharp/issues/824
+                int widthInBlocks = Math.Min(this.Component.SpectralBlocks.Width, this.SizeInBlocks.Width);
 
-                for (int xBlock = 0; xBlock < this.SizeInBlocks.Width; xBlock++)
+                for (int xBlock = 0; xBlock < widthInBlocks; xBlock++)
                 {
-                    ref Block8x8 block = ref Unsafe.Add(ref blockRowBase, xBlock);
+                    ref Block8x8 block = ref blockRow[xBlock];
                     int xBuffer = xBlock * this.blockAreaSize.Width;
+                    ref float destAreaOrigin = ref colorBufferRow[xBuffer];
 
-                    BufferArea<float> destArea = this.ColorBuffer.GetArea(
-                        xBuffer,
-                        yBuffer,
-                        this.blockAreaSize.Width,
-                        this.blockAreaSize.Height);
-
-                    blockPp.ProcessBlockColorsInto(ref block, destArea, maximumValue);
+                    blockPp.ProcessBlockColorsInto(ref block, ref destAreaOrigin, destAreaStride, maximumValue);
                 }
             }
 

src/ImageSharp/Memory/Allocators/ArrayPoolMemoryAllocator.Buffer{T}.cs

@@ -3,6 +3,7 @@
 
 using System;
 using System.Buffers;
+using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 using SixLabors.ImageSharp.Memory.Internals;
 
@@ -50,7 +51,7 @@ namespace SixLabors.ImageSharp.Memory
             {
                 if (this.Data is null)
                 {
-                    throw new ObjectDisposedException("ArrayPoolMemoryAllocator.Buffer<T>");
+                    ThrowObjectDisposedException();
                 }
 
                 return MemoryMarshal.Cast<byte, T>(this.Data.AsSpan()).Slice(0, this.length);
@@ -74,6 +75,12 @@ namespace SixLabors.ImageSharp.Memory
             }
 
             protected override object GetPinnableObject() => this.Data;
+
+            [MethodImpl(InliningOptions.ColdPath)]
+            private static void ThrowObjectDisposedException()
+            {
+                throw new ObjectDisposedException("ArrayPoolMemoryAllocator.Buffer<T>");
+            }
         }
 
         /// <summary>

src/ImageSharp/Memory/Allocators/ArrayPoolMemoryAllocator.cs

@@ -133,8 +133,7 @@ namespace SixLabors.ImageSharp.Memory
             int bufferSizeInBytes = length * itemSizeBytes;
             if (bufferSizeInBytes < 0 || bufferSizeInBytes > this.BufferCapacityInBytes)
             {
-                throw new InvalidMemoryOperationException(
-                    $"Requested allocation: {length} elements of {typeof(T).Name} is over the capacity of the MemoryAllocator.");
+                ThrowInvalidAllocationException<T>(length);
             }
 
             ArrayPool<byte> pool = this.GetArrayPool(bufferSizeInBytes);
@@ -171,6 +170,11 @@ namespace SixLabors.ImageSharp.Memory
             return maxPoolSizeInBytes / 4;
         }
 
+        [MethodImpl(InliningOptions.ColdPath)]
+        private static void ThrowInvalidAllocationException<T>(int length) =>
+            throw new InvalidMemoryOperationException(
+                $"Requested allocation: {length} elements of {typeof(T).Name} is over the capacity of the MemoryAllocator.");
+
         private ArrayPool<byte> GetArrayPool(int bufferSizeInBytes)
         {
             return bufferSizeInBytes <= this.PoolSelectorThresholdInBytes ? this.normalArrayPool : this.largeArrayPool;

src/ImageSharp/Memory/Buffer2DExtensions.cs

@@ -27,52 +27,6 @@ namespace SixLabors.ImageSharp.Memory
             return buffer.FastMemoryGroup.View;
         }
 
-        /// <summary>
-        /// Gets a <see cref="Span{T}"/> to the backing data of <paramref name="buffer"/>
-        /// if the backing group consists of one single contiguous memory buffer.
-        /// Throws <see cref="InvalidOperationException"/> otherwise.
-        /// </summary>
-        /// <param name="buffer">The <see cref="Buffer2D{T}"/>.</param>
-        /// <typeparam name="T">The value type.</typeparam>
-        /// <returns>The <see cref="Span{T}"/> referencing the memory area.</returns>
-        /// <exception cref="InvalidOperationException">
-        /// Thrown when the backing group is discontiguous.
-        /// </exception>
-        internal static Span<T> GetSingleSpan<T>(this Buffer2D<T> buffer)
-            where T : struct
-        {
-            Guard.NotNull(buffer, nameof(buffer));
-            if (buffer.FastMemoryGroup.Count > 1)
-            {
-                throw new InvalidOperationException("GetSingleSpan is only valid for a single-buffer group!");
-            }
-
-            return buffer.FastMemoryGroup.Single().Span;
-        }
-
-        /// <summary>
-        /// Gets a <see cref="Memory{T}"/> to the backing data of <paramref name="buffer"/>
-        /// if the backing group consists of one single contiguous memory buffer.
-        /// Throws <see cref="InvalidOperationException"/> otherwise.
-        /// </summary>
-        /// <param name="buffer">The <see cref="Buffer2D{T}"/>.</param>
-        /// <typeparam name="T">The value type.</typeparam>
-        /// <returns>The <see cref="Memory{T}"/>.</returns>
-        /// <exception cref="InvalidOperationException">
-        /// Thrown when the backing group is discontiguous.
-        /// </exception>
-        internal static Memory<T> GetSingleMemory<T>(this Buffer2D<T> buffer)
-            where T : struct
-        {
-            Guard.NotNull(buffer, nameof(buffer));
-            if (buffer.FastMemoryGroup.Count > 1)
-            {
-                throw new InvalidOperationException("GetSingleMemory is only valid for a single-buffer group!");
-            }
-
-            return buffer.FastMemoryGroup.Single();
-        }
-
         /// <summary>
         /// TODO: Does not work with multi-buffer groups, should be specific to Resize.
         /// Copy <paramref name="columnCount"/> columns of <paramref name="buffer"/> inplace,

src/ImageSharp/Memory/Buffer2D{T}.cs

@@ -2,6 +2,7 @@
 // Licensed under the Apache License, Version 2.0.
 
 using System;
+using System.Linq;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 
@@ -158,6 +159,36 @@ namespace SixLabors.ImageSharp.Memory
             return this.FastMemoryGroup.View.GetBoundedSlice(y * this.Width, this.Width);
         }
 
+        /// <summary>
+        /// Gets a <see cref="Span{T}"/> to the backing data if the backing group consists of a single contiguous memory buffer.
+        /// Throws <see cref="InvalidOperationException"/> otherwise.
+        /// </summary>
+        /// <returns>The <see cref="Span{T}"/> referencing the memory area.</returns>
+        /// <exception cref="InvalidOperationException">
+        /// Thrown when the backing group is discontiguous.
+        /// </exception>
+        [MethodImpl(InliningOptions.ShortMethod)]
+        internal Span<T> GetSingleSpan()
+        {
+            // TODO: If we need a public version of this method, we need to cache the non-fast Memory<T> of this.MemoryGroup
+            return this.cachedMemory.Length != 0 ? this.cachedMemory.Span : this.GetSingleSpanSlow();
+        }
+
+        /// <summary>
+        /// Gets a <see cref="Memory{T}"/> to the backing data of if the backing group consists of a single contiguous memory buffer.
+        /// Throws <see cref="InvalidOperationException"/> otherwise.
+        /// </summary>
+        /// <returns>The <see cref="Memory{T}"/>.</returns>
+        /// <exception cref="InvalidOperationException">
+        /// Thrown when the backing group is discontiguous.
+        /// </exception>
+        [MethodImpl(InliningOptions.ShortMethod)]
+        internal Memory<T> GetSingleMemory()
+        {
+            // TODO: If we need a public version of this method, we need to cache the non-fast Memory<T> of this.MemoryGroup
+            return this.cachedMemory.Length != 0 ? this.cachedMemory : this.GetSingleMemorySlow();
+        }
+
         /// <summary>
         /// Swaps the contents of 'destination' with 'source' if the buffers are owned (1),
         /// copies the contents of 'source' to 'destination' otherwise (2). Buffers should be of same size in case 2!
@@ -171,6 +202,12 @@ namespace SixLabors.ImageSharp.Memory
         [MethodImpl(InliningOptions.ColdPath)]
         private Memory<T> GetRowMemorySlow(int y) => this.FastMemoryGroup.GetBoundedSlice(y * this.Width, this.Width);
 
+        [MethodImpl(InliningOptions.ColdPath)]
+        private Memory<T> GetSingleMemorySlow() => this.FastMemoryGroup.Single();
+
+        [MethodImpl(InliningOptions.ColdPath)]
+        private Span<T> GetSingleSpanSlow() => this.FastMemoryGroup.Single().Span;
+
         [MethodImpl(InliningOptions.ColdPath)]
         private ref T GetElementSlow(int x, int y)
         {

src/ImageSharp/Memory/BufferArea{T}.cs

@@ -79,8 +79,12 @@ namespace SixLabors.ImageSharp.Memory
         /// </summary>
         /// <returns>The reference to the [0,0] element</returns>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public ref T GetReferenceToOrigin() =>
-            ref this.GetRowSpan(0)[0];
+        public ref T GetReferenceToOrigin()
+        {
+            int y = this.Rectangle.Y;
+            int x = this.Rectangle.X;
+            return ref this.DestinationBuffer.GetRowSpan(y)[x];
+        }
 
         /// <summary>
         /// Gets a span to row 'y' inside this area.
@@ -90,11 +94,11 @@ namespace SixLabors.ImageSharp.Memory
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public Span<T> GetRowSpan(int y)
         {
-            int yy = this.GetRowIndex(y);
+            int yy = this.Rectangle.Y + y;
             int xx = this.Rectangle.X;
             int width = this.Rectangle.Width;
 
-            return this.DestinationBuffer.FastMemoryGroup.GetBoundedSlice(yy + xx, width).Span;
+            return this.DestinationBuffer.GetRowSpan(yy).Slice(xx, width);
         }
 
         /// <summary>
@@ -129,12 +133,6 @@ namespace SixLabors.ImageSharp.Memory
             return new BufferArea<T>(this.DestinationBuffer, rectangle);
         }
 
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        internal int GetRowIndex(int y)
-        {
-            return (y + this.Rectangle.Y) * this.DestinationBuffer.Width;
-        }
-
         public void Clear()
         {
             // Optimization for when the size of the area is the same as the buffer size.

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

@@ -29,7 +29,7 @@ namespace SixLabors.ImageSharp.PixelFormats
                 Guard.DestinationShouldNotBeTooShort(sourcePixels, destinationVectors, nameof(destinationVectors));
 
                 destinationVectors = destinationVectors.Slice(0, sourcePixels.Length);
-                SimdUtils.BulkConvertByteToNormalizedFloat(
+                SimdUtils.ByteToNormalizedFloat(
                     MemoryMarshal.Cast<Rgba32, byte>(sourcePixels),
                     MemoryMarshal.Cast<Vector4, float>(destinationVectors));
                 Vector4Converters.ApplyForwardConversionModifiers(destinationVectors, modifiers);
@@ -46,7 +46,7 @@ namespace SixLabors.ImageSharp.PixelFormats
 
                 destinationPixels = destinationPixels.Slice(0, sourceVectors.Length);
                 Vector4Converters.ApplyBackwardConversionModifiers(sourceVectors, modifiers);
-                SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows(
+                SimdUtils.NormalizedFloatToByteSaturate(
                     MemoryMarshal.Cast<Vector4, float>(sourceVectors),
                     MemoryMarshal.Cast<Rgba32, byte>(destinationPixels));
             }

src/ImageSharp/PixelFormats/Utils/Vector4Converters.RgbaCompatible.cs

@@ -62,7 +62,7 @@ namespace SixLabors.ImageSharp.PixelFormats.Utils
 
                 // 'destVectors' and 'lastQuarterOfDestBuffer' are overlapping buffers,
                 // but we are always reading/writing at different positions:
-                SimdUtils.BulkConvertByteToNormalizedFloat(
+                SimdUtils.ByteToNormalizedFloat(
                     MemoryMarshal.Cast<Rgba32, byte>(lastQuarterOfDestBuffer),
                     MemoryMarshal.Cast<Vector4, float>(destVectors.Slice(0, countWithoutLastItem)));
 
@@ -107,7 +107,7 @@ namespace SixLabors.ImageSharp.PixelFormats.Utils
                 {
                     Span<Rgba32> tempSpan = tempBuffer.Memory.Span;
 
-                    SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows(
+                    SimdUtils.NormalizedFloatToByteSaturate(
                         MemoryMarshal.Cast<Vector4, float>(sourceVectors),
                         MemoryMarshal.Cast<Rgba32, byte>(tempSpan));
 
@@ -122,7 +122,7 @@ namespace SixLabors.ImageSharp.PixelFormats.Utils
                     return int.MaxValue;
                 }
 
-                return SimdUtils.ExtendedIntrinsics.IsAvailable && SimdUtils.IsAvx2CompatibleArchitecture ? 256 : 128;
+                return SimdUtils.ExtendedIntrinsics.IsAvailable && SimdUtils.HasVector8 ? 256 : 128;
             }
         }
     }

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

@@ -3,6 +3,7 @@
 
 using System;
 using System.Buffers;
+using System.Diagnostics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 using SixLabors.ImageSharp.Memory;
@@ -249,12 +250,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Transforms
         private unsafe ResizeKernel CreateKernel(int dataRowIndex, int left, int right)
         {
             int length = right - left + 1;
-
-            if (length > this.data.Width)
-            {
-                throw new InvalidOperationException(
-                    $"Error in KernelMap.CreateKernel({dataRowIndex},{left},{right}): left > this.data.Width");
-            }
+            this.ValidateSizesForCreateKernel(length, dataRowIndex, left, right);
 
             Span<float> rowSpan = this.data.GetRowSpan(dataRowIndex);
 
@@ -262,5 +258,15 @@ namespace SixLabors.ImageSharp.Processing.Processors.Transforms
             float* rowPtr = (float*)Unsafe.AsPointer(ref rowReference);
             return new ResizeKernel(left, rowPtr, length);
         }
+
+        [Conditional("DEBUG")]
+        private void ValidateSizesForCreateKernel(int length, int dataRowIndex, int left, int right)
+        {
+            if (length > this.data.Width)
+            {
+                throw new InvalidOperationException(
+                    $"Error in KernelMap.CreateKernel({dataRowIndex},{left},{right}): left > this.data.Width");
+            }
+        }
     }
 }

tests/ImageSharp.Benchmarks/Codecs/Jpeg/BlockOperations/Block8x8F_CopyTo1x1.cs

@@ -4,41 +4,66 @@
 using System;
 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
 
 using BenchmarkDotNet.Attributes;
 
 using SixLabors.ImageSharp.Formats.Jpeg.Components;
-using SixLabors.ImageSharp.Memory;
 
 // ReSharper disable InconsistentNaming
 namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg.BlockOperations
 {
-    public class Block8x8F_CopyTo1x1
+    public unsafe class Block8x8F_CopyTo1x1
     {
         private Block8x8F block;
+        private readonly Block8x8F[] blockArray = new Block8x8F[1];
 
-        private Buffer2D<float> buffer;
+        private static readonly int Width = 100;
 
-        private BufferArea<float> destArea;
+        private float[] buffer = new float[Width * 500];
+        private readonly float[] unpinnedBuffer = new float[Width * 500];
+        private GCHandle bufferHandle;
+        private GCHandle blockHandle;
+        private float* bufferPtr;
+        private float* blockPtr;
 
         [GlobalSetup]
         public void Setup()
         {
-            if (!SimdUtils.IsAvx2CompatibleArchitecture)
+            if (!SimdUtils.HasVector8)
             {
                 throw new InvalidOperationException("Benchmark Block8x8F_CopyTo1x1 is invalid on platforms without AVX2 support.");
             }
 
-            this.buffer = Configuration.Default.MemoryAllocator.Allocate2D<float>(1000, 500);
-            this.destArea = this.buffer.GetArea(200, 100, 64, 64);
+            this.bufferHandle = GCHandle.Alloc(this.buffer, GCHandleType.Pinned);
+            this.bufferPtr = (float*)this.bufferHandle.AddrOfPinnedObject();
+
+            // Pin self so we can take address of to the block:
+            this.blockHandle = GCHandle.Alloc(this.blockArray, GCHandleType.Pinned);
+            this.blockPtr = (float*)Unsafe.AsPointer(ref this.block);
+        }
+
+        [GlobalCleanup]
+        public void Cleanup()
+        {
+            this.bufferPtr = null;
+            this.blockPtr = null;
+            this.bufferHandle.Free();
+            this.blockHandle.Free();
+            this.buffer = null;
         }
 
         [Benchmark(Baseline = true)]
         public void Original()
         {
             ref byte selfBase = ref Unsafe.As<Block8x8F, byte>(ref this.block);
-            ref byte destBase = ref Unsafe.As<float, byte>(ref this.destArea.GetReferenceToOrigin());
-            int destStride = this.destArea.Stride * sizeof(float);
+            ref byte destBase = ref Unsafe.AsRef<byte>(this.bufferPtr);
+            int destStride = Width * sizeof(float);
 
             CopyRowImpl(ref selfBase, ref destBase, destStride, 0);
             CopyRowImpl(ref selfBase, ref destBase, destStride, 1);
@@ -58,12 +83,12 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg.BlockOperations
             Unsafe.CopyBlock(ref d, ref s, 8 * sizeof(float));
         }
 
-        [Benchmark]
+        // [Benchmark]
         public void UseVector8()
         {
             ref Block8x8F s = ref this.block;
-            ref float origin = ref this.destArea.GetReferenceToOrigin();
-            int stride = this.destArea.Stride;
+            ref float origin = ref Unsafe.AsRef<float>(this.bufferPtr);
+            int stride = Width;
 
             ref Vector<float> d0 = ref Unsafe.As<float, Vector<float>>(ref origin);
             ref Vector<float> d1 = ref Unsafe.As<float, Vector<float>>(ref Unsafe.Add(ref origin, stride));
@@ -93,12 +118,12 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg.BlockOperations
             d7 = row7;
         }
 
-        [Benchmark]
+        // [Benchmark]
         public void UseVector8_V2()
         {
             ref Block8x8F s = ref this.block;
-            ref float origin = ref this.destArea.GetReferenceToOrigin();
-            int stride = this.destArea.Stride;
+            ref float origin = ref Unsafe.AsRef<float>(this.bufferPtr);
+            int stride = Width;
 
             ref Vector<float> d0 = ref Unsafe.As<float, Vector<float>>(ref origin);
             ref Vector<float> d1 = ref Unsafe.As<float, Vector<float>>(ref Unsafe.Add(ref origin, stride));
@@ -119,15 +144,247 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg.BlockOperations
             d7 = Unsafe.As<Vector4, Vector<float>>(ref s.V7L);
         }
 
-        // RESULTS:
+        [Benchmark]
+        public void UseVector8_V3()
+        {
+            int stride = Width * sizeof(float);
+            ref float d = ref this.unpinnedBuffer[0];
+            ref Vector<float> s = ref Unsafe.As<Block8x8F, Vector<float>>(ref this.block);
+
+            Vector<float> v0 = s;
+            Vector<float> v1 = Unsafe.AddByteOffset(ref s, (IntPtr)1);
+            Vector<float> v2 = Unsafe.AddByteOffset(ref s, (IntPtr)2);
+            Vector<float> v3 = Unsafe.AddByteOffset(ref s, (IntPtr)3);
+
+            Unsafe.As<float, Vector<float>>(ref d) = v0;
+            Unsafe.As<float, Vector<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)stride)) = v1;
+            Unsafe.As<float, Vector<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 2))) = v2;
+            Unsafe.As<float, Vector<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 3))) = v3;
+
+            v0 = Unsafe.AddByteOffset(ref s, (IntPtr)4);
+            v1 = Unsafe.AddByteOffset(ref s, (IntPtr)5);
+            v2 = Unsafe.AddByteOffset(ref s, (IntPtr)6);
+            v3 = Unsafe.AddByteOffset(ref s, (IntPtr)7);
+
+            Unsafe.As<float, Vector<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 4))) = v0;
+            Unsafe.As<float, Vector<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 5))) = v1;
+            Unsafe.As<float, Vector<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 6))) = v2;
+            Unsafe.As<float, Vector<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 7))) = v3;
+        }
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+        [Benchmark]
+        public void UseVector256_Avx2_Variant1()
+        {
+            int stride = Width;
+            float* d = this.bufferPtr;
+            float* s = this.blockPtr;
+            Vector256<float> v;
+
+            v = Avx.LoadVector256(s);
+            Avx.Store(d, v);
+
+            v = Avx.LoadVector256(s + 8);
+            Avx.Store(d + stride, v);
+
+            v = Avx.LoadVector256(s + (8 * 2));
+            Avx.Store(d + (stride * 2), v);
+
+            v = Avx.LoadVector256(s + (8 * 3));
+            Avx.Store(d + (stride * 3), v);
+
+            v = Avx.LoadVector256(s + (8 * 4));
+            Avx.Store(d + (stride * 4), v);
+
+            v = Avx.LoadVector256(s + (8 * 5));
+            Avx.Store(d + (stride * 5), v);
+
+            v = Avx.LoadVector256(s + (8 * 6));
+            Avx.Store(d + (stride * 6), v);
+
+            v = Avx.LoadVector256(s + (8 * 7));
+            Avx.Store(d + (stride * 7), v);
+        }
+
+        [Benchmark]
+        public void UseVector256_Avx2_Variant2()
+        {
+            int stride = Width;
+            float* d = this.bufferPtr;
+            float* s = this.blockPtr;
+
+            Vector256<float> v0 = Avx.LoadVector256(s);
+            Vector256<float> v1 = Avx.LoadVector256(s + 8);
+            Vector256<float> v2 = Avx.LoadVector256(s + (8 * 2));
+            Vector256<float> v3 = Avx.LoadVector256(s + (8 * 3));
+            Vector256<float> v4 = Avx.LoadVector256(s + (8 * 4));
+            Vector256<float> v5 = Avx.LoadVector256(s + (8 * 5));
+            Vector256<float> v6 = Avx.LoadVector256(s + (8 * 6));
+            Vector256<float> v7 = Avx.LoadVector256(s + (8 * 7));
+
+            Avx.Store(d, v0);
+            Avx.Store(d + stride, v1);
+            Avx.Store(d + (stride * 2), v2);
+            Avx.Store(d + (stride * 3), v3);
+            Avx.Store(d + (stride * 4), v4);
+            Avx.Store(d + (stride * 5), v5);
+            Avx.Store(d + (stride * 6), v6);
+            Avx.Store(d + (stride * 7), v7);
+        }
+
+        [Benchmark]
+        public void UseVector256_Avx2_Variant3()
+        {
+            int stride = Width;
+            float* d = this.bufferPtr;
+            float* s = this.blockPtr;
+
+            Vector256<float> v0 = Avx.LoadVector256(s);
+            Vector256<float> v1 = Avx.LoadVector256(s + 8);
+            Vector256<float> v2 = Avx.LoadVector256(s + (8 * 2));
+            Vector256<float> v3 = Avx.LoadVector256(s + (8 * 3));
+            Avx.Store(d, v0);
+            Avx.Store(d + stride, v1);
+            Avx.Store(d + (stride * 2), v2);
+            Avx.Store(d + (stride * 3), v3);
+
+            v0 = Avx.LoadVector256(s + (8 * 4));
+            v1 = Avx.LoadVector256(s + (8 * 5));
+            v2 = Avx.LoadVector256(s + (8 * 6));
+            v3 = Avx.LoadVector256(s + (8 * 7));
+            Avx.Store(d + (stride * 4), v0);
+            Avx.Store(d + (stride * 5), v1);
+            Avx.Store(d + (stride * 6), v2);
+            Avx.Store(d + (stride * 7), v3);
+        }
+
+        [Benchmark]
+        public void UseVector256_Avx2_Variant3_RefCast()
+        {
+            int stride = Width;
+            ref float d = ref this.unpinnedBuffer[0];
+            ref Vector256<float> s = ref Unsafe.As<Block8x8F, Vector256<float>>(ref this.block);
+
+            Vector256<float> v0 = s;
+            Vector256<float> v1 = Unsafe.Add(ref s, 1);
+            Vector256<float> v2 = Unsafe.Add(ref s, 2);
+            Vector256<float> v3 = Unsafe.Add(ref s, 3);
+
+            Unsafe.As<float, Vector256<float>>(ref d) = v0;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.Add(ref d, stride)) = v1;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.Add(ref d, stride * 2)) = v2;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.Add(ref d, stride * 3)) = v3;
+
+            v0 = Unsafe.Add(ref s, 4);
+            v1 = Unsafe.Add(ref s, 5);
+            v2 = Unsafe.Add(ref s, 6);
+            v3 = Unsafe.Add(ref s, 7);
+
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.Add(ref d, stride * 4)) = v0;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.Add(ref d, stride * 5)) = v1;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.Add(ref d, stride * 6)) = v2;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.Add(ref d, stride * 7)) = v3;
+        }
+
+        [Benchmark]
+        public void UseVector256_Avx2_Variant3_RefCast_Mod()
+        {
+            int stride = Width * sizeof(float);
+            ref float d = ref this.unpinnedBuffer[0];
+            ref Vector256<float> s = ref Unsafe.As<Block8x8F, Vector256<float>>(ref this.block);
+
+            Vector256<float> v0 = s;
+            Vector256<float> v1 = Unsafe.AddByteOffset(ref s, (IntPtr)1);
+            Vector256<float> v2 = Unsafe.AddByteOffset(ref s, (IntPtr)2);
+            Vector256<float> v3 = Unsafe.AddByteOffset(ref s, (IntPtr)3);
+
+            Unsafe.As<float, Vector256<float>>(ref d) = v0;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)stride)) = v1;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 2))) = v2;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 3))) = v3;
+
+            v0 = Unsafe.AddByteOffset(ref s, (IntPtr)4);
+            v1 = Unsafe.AddByteOffset(ref s, (IntPtr)5);
+            v2 = Unsafe.AddByteOffset(ref s, (IntPtr)6);
+            v3 = Unsafe.AddByteOffset(ref s, (IntPtr)7);
+
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 4))) = v0;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 5))) = v1;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 6))) = v2;
+            Unsafe.As<float, Vector256<float>>(ref Unsafe.AddByteOffset(ref d, (IntPtr)(stride * 7))) = v3;
+        }
+
+        // [Benchmark]
+        public void UseVector256_Avx2_Variant3_WithLocalPinning()
+        {
+            int stride = Width;
+            fixed (float* d = this.unpinnedBuffer)
+            fixed (Block8x8F* ss = &this.block)
+            {
+                var s = (float*)ss;
+                Vector256<float> v0 = Avx.LoadVector256(s);
+                Vector256<float> v1 = Avx.LoadVector256(s + 8);
+                Vector256<float> v2 = Avx.LoadVector256(s + (8 * 2));
+                Vector256<float> v3 = Avx.LoadVector256(s + (8 * 3));
+                Avx.Store(d, v0);
+                Avx.Store(d + stride, v1);
+                Avx.Store(d + (stride * 2), v2);
+                Avx.Store(d + (stride * 3), v3);
+
+                v0 = Avx.LoadVector256(s + (8 * 4));
+                v1 = Avx.LoadVector256(s + (8 * 5));
+                v2 = Avx.LoadVector256(s + (8 * 6));
+                v3 = Avx.LoadVector256(s + (8 * 7));
+                Avx.Store(d + (stride * 4), v0);
+                Avx.Store(d + (stride * 5), v1);
+                Avx.Store(d + (stride * 6), v2);
+                Avx.Store(d + (stride * 7), v3);
+            }
+        }
+
+        // [Benchmark]
+        public void UseVector256_Avx2_Variant3_sbyte()
+        {
+            int stride = Width * 4;
+            var d = (sbyte*)this.bufferPtr;
+            var s = (sbyte*)this.blockPtr;
+
+            Vector256<sbyte> v0 = Avx.LoadVector256(s);
+            Vector256<sbyte> v1 = Avx.LoadVector256(s + 32);
+            Vector256<sbyte> v2 = Avx.LoadVector256(s + (32 * 2));
+            Vector256<sbyte> v3 = Avx.LoadVector256(s + (32 * 3));
+            Avx.Store(d, v0);
+            Avx.Store(d + stride, v1);
+            Avx.Store(d + (stride * 2), v2);
+            Avx.Store(d + (stride * 3), v3);
+
+            v0 = Avx.LoadVector256(s + (32 * 4));
+            v1 = Avx.LoadVector256(s + (32 * 5));
+            v2 = Avx.LoadVector256(s + (32 * 6));
+            v3 = Avx.LoadVector256(s + (32 * 7));
+            Avx.Store(d + (stride * 4), v0);
+            Avx.Store(d + (stride * 5), v1);
+            Avx.Store(d + (stride * 6), v2);
+            Avx.Store(d + (stride * 7), v3);
+        }
+#endif
+
+        // *** RESULTS 02/2020 ***
+        // BenchmarkDotNet=v0.12.0, OS=Windows 10.0.18363
+        // Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores
+        // .NET Core SDK=3.1.200-preview-014971
+        //   [Host]     : .NET Core 3.1.2 (CoreCLR 4.700.20.6602, CoreFX 4.700.20.6702), X64 RyuJIT
+        //   DefaultJob : .NET Core 3.1.2 (CoreCLR 4.700.20.6602, CoreFX 4.700.20.6702), X64 RyuJIT
         //
-        //         Method |     Mean |     Error |    StdDev | Scaled |
-        // -------------- |---------:|----------:|----------:|-------:|
-        //       Original | 22.53 ns | 0.1660 ns | 0.1553 ns |   1.00 |
-        //     UseVector8 | 21.59 ns | 0.3079 ns | 0.2571 ns |   0.96 |
-        //  UseVector8_V2 | 22.57 ns | 0.1699 ns | 0.1506 ns |   1.00 |
         //
-        // Conclusion:
-        // Doesn't worth to bother with this
+        // |                                 Method |     Mean |     Error |    StdDev | Ratio | RatioSD |
+        // |--------------------------------------- |---------:|----------:|----------:|------:|--------:|
+        // |                               Original | 4.012 ns | 0.0567 ns | 0.0531 ns |  1.00 |    0.00 |
+        // |                          UseVector8_V3 | 4.013 ns | 0.0947 ns | 0.0840 ns |  1.00 |    0.03 |
+        // |             UseVector256_Avx2_Variant1 | 2.546 ns | 0.0376 ns | 0.0314 ns |  0.63 |    0.01 |
+        // |             UseVector256_Avx2_Variant2 | 2.643 ns | 0.0162 ns | 0.0151 ns |  0.66 |    0.01 |
+        // |             UseVector256_Avx2_Variant3 | 2.520 ns | 0.0760 ns | 0.0813 ns |  0.63 |    0.02 |
+        // |     UseVector256_Avx2_Variant3_RefCast | 2.300 ns | 0.0877 ns | 0.0938 ns |  0.58 |    0.03 |
+        // | UseVector256_Avx2_Variant3_RefCast_Mod | 2.139 ns | 0.0698 ns | 0.0686 ns |  0.53 |    0.02 |
     }
 }

tests/ImageSharp.Benchmarks/Codecs/Jpeg/BlockOperations/Block8x8F_Round.cs

@@ -4,6 +4,12 @@
 using System;
 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
 
 using BenchmarkDotNet.Attributes;
 
@@ -12,10 +18,14 @@ using SixLabors.ImageSharp.Formats.Jpeg.Components;
 // ReSharper disable InconsistentNaming
 namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg.BlockOperations
 {
-    public class Block8x8F_Round
+    public unsafe class Block8x8F_Round
     {
         private Block8x8F block;
 
+        private readonly byte[] blockBuffer = new byte[512];
+        private GCHandle blockHandle;
+        private float* alignedPtr;
+
         [GlobalSetup]
         public void Setup()
         {
@@ -24,13 +34,27 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg.BlockOperations
                 throw new NotSupportedException("Vector<float>.Count != 8");
             }
 
-            for (int i = 0; i < Block8x8F.Size; i++)
+            this.blockHandle = GCHandle.Alloc(this.blockBuffer, GCHandleType.Pinned);
+            ulong ptr = (ulong)this.blockHandle.AddrOfPinnedObject();
+            ptr += 16;
+            ptr -= ptr % 16;
+
+            if (ptr % 16 != 0)
             {
-                this.block[i] = i * 44.8f;
+                throw new Exception("ptr is unaligned");
             }
+
+            this.alignedPtr = (float*)ptr;
         }
 
-        [Benchmark(Baseline = true)]
+        [GlobalCleanup]
+        public void Cleanup()
+        {
+            this.blockHandle.Free();
+            this.alignedPtr = null;
+        }
+
+        [Benchmark]
         public void ScalarRound()
         {
             ref float b = ref Unsafe.As<Block8x8F, float>(ref this.block);
@@ -42,8 +66,8 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg.BlockOperations
             }
         }
 
-        [Benchmark]
-        public void SimdRound()
+        [Benchmark(Baseline = true)]
+        public void SimdUtils_FastRound_Vector8()
         {
             ref Block8x8F b = ref this.block;
 
@@ -64,5 +88,411 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg.BlockOperations
             ref Vector<float> row7 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V7L);
             row7 = SimdUtils.FastRound(row7);
         }
+
+        [Benchmark]
+        public void SimdUtils_FastRound_Vector8_ForceAligned()
+        {
+            ref Block8x8F b = ref Unsafe.AsRef<Block8x8F>(this.alignedPtr);
+
+            ref Vector<float> row0 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V0L);
+            row0 = SimdUtils.FastRound(row0);
+            ref Vector<float> row1 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V1L);
+            row1 = SimdUtils.FastRound(row1);
+            ref Vector<float> row2 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V2L);
+            row2 = SimdUtils.FastRound(row2);
+            ref Vector<float> row3 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V3L);
+            row3 = SimdUtils.FastRound(row3);
+            ref Vector<float> row4 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V4L);
+            row4 = SimdUtils.FastRound(row4);
+            ref Vector<float> row5 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V5L);
+            row5 = SimdUtils.FastRound(row5);
+            ref Vector<float> row6 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V6L);
+            row6 = SimdUtils.FastRound(row6);
+            ref Vector<float> row7 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V7L);
+            row7 = SimdUtils.FastRound(row7);
+        }
+
+        [Benchmark]
+        public void SimdUtils_FastRound_Vector8_Grouped()
+        {
+            ref Block8x8F b = ref this.block;
+
+            ref Vector<float> row0 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V0L);
+            ref Vector<float> row1 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V1L);
+            ref Vector<float> row2 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V2L);
+            ref Vector<float> row3 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V3L);
+
+            row0 = SimdUtils.FastRound(row0);
+            row1 = SimdUtils.FastRound(row1);
+            row2 = SimdUtils.FastRound(row2);
+            row3 = SimdUtils.FastRound(row3);
+
+            row0 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V4L);
+            row1 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V5L);
+            row2 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V6L);
+            row3 = ref Unsafe.As<Vector4, Vector<float>>(ref b.V7L);
+
+            row0 = SimdUtils.FastRound(row0);
+            row1 = SimdUtils.FastRound(row1);
+            row2 = SimdUtils.FastRound(row2);
+            row3 = SimdUtils.FastRound(row3);
+        }
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+        [Benchmark]
+        public void Sse41_V1()
+        {
+            ref Vector128<float> b0 = ref Unsafe.As<Block8x8F, Vector128<float>>(ref this.block);
+
+            ref Vector128<float> p = ref b0;
+            p = Sse41.RoundToNearestInteger(p);
+
+            p = ref Unsafe.Add(ref b0, 1);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 2);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 3);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 4);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 5);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 6);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 7);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 8);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 9);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 10);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 11);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 12);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 13);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 14);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.Add(ref b0, 15);
+            p = Sse41.RoundToNearestInteger(p);
+        }
+
+        [Benchmark]
+        public unsafe void Sse41_V2()
+        {
+            ref Vector128<float> p = ref Unsafe.As<Block8x8F, Vector128<float>>(ref this.block);
+            p = Sse41.RoundToNearestInteger(p);
+            var offset = (IntPtr)sizeof(Vector128<float>);
+            p = Sse41.RoundToNearestInteger(p);
+
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+            p = ref Unsafe.AddByteOffset(ref p, offset);
+            p = Sse41.RoundToNearestInteger(p);
+        }
+
+        [Benchmark]
+        public unsafe void Sse41_V3()
+        {
+            ref Vector128<float> p = ref Unsafe.As<Block8x8F, Vector128<float>>(ref this.block);
+            p = Sse41.RoundToNearestInteger(p);
+            var offset = (IntPtr)sizeof(Vector128<float>);
+
+            for (int i = 0; i < 15; i++)
+            {
+                p = ref Unsafe.AddByteOffset(ref p, offset);
+                p = Sse41.RoundToNearestInteger(p);
+            }
+        }
+
+        [Benchmark]
+        public unsafe void Sse41_V4()
+        {
+            ref Vector128<float> p = ref Unsafe.As<Block8x8F, Vector128<float>>(ref this.block);
+            var offset = (IntPtr)sizeof(Vector128<float>);
+
+            ref Vector128<float> a = ref p;
+            ref Vector128<float> b = ref Unsafe.AddByteOffset(ref a, offset);
+            ref Vector128<float> c = ref Unsafe.AddByteOffset(ref b, offset);
+            ref Vector128<float> d = ref Unsafe.AddByteOffset(ref c, offset);
+            a = Sse41.RoundToNearestInteger(a);
+            b = Sse41.RoundToNearestInteger(b);
+            c = Sse41.RoundToNearestInteger(c);
+            d = Sse41.RoundToNearestInteger(d);
+
+            a = ref Unsafe.AddByteOffset(ref d, offset);
+            b = ref Unsafe.AddByteOffset(ref a, offset);
+            c = ref Unsafe.AddByteOffset(ref b, offset);
+            d = ref Unsafe.AddByteOffset(ref c, offset);
+            a = Sse41.RoundToNearestInteger(a);
+            b = Sse41.RoundToNearestInteger(b);
+            c = Sse41.RoundToNearestInteger(c);
+            d = Sse41.RoundToNearestInteger(d);
+
+            a = ref Unsafe.AddByteOffset(ref d, offset);
+            b = ref Unsafe.AddByteOffset(ref a, offset);
+            c = ref Unsafe.AddByteOffset(ref b, offset);
+            d = ref Unsafe.AddByteOffset(ref c, offset);
+            a = Sse41.RoundToNearestInteger(a);
+            b = Sse41.RoundToNearestInteger(b);
+            c = Sse41.RoundToNearestInteger(c);
+            d = Sse41.RoundToNearestInteger(d);
+
+            a = ref Unsafe.AddByteOffset(ref d, offset);
+            b = ref Unsafe.AddByteOffset(ref a, offset);
+            c = ref Unsafe.AddByteOffset(ref b, offset);
+            d = ref Unsafe.AddByteOffset(ref c, offset);
+            a = Sse41.RoundToNearestInteger(a);
+            b = Sse41.RoundToNearestInteger(b);
+            c = Sse41.RoundToNearestInteger(c);
+            d = Sse41.RoundToNearestInteger(d);
+        }
+
+        [Benchmark]
+        public unsafe void Sse41_V5_Unaligned()
+        {
+            float* p = this.alignedPtr + 1;
+
+            Vector128<float> v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+            p += 8;
+
+            v = Sse.LoadVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.Store(p, v);
+        }
+
+        [Benchmark]
+        public unsafe void Sse41_V5_Aligned()
+        {
+            float* p = this.alignedPtr;
+
+            Vector128<float> v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+
+            v = Sse.LoadAlignedVector128(p);
+            v = Sse41.RoundToNearestInteger(v);
+            Sse.StoreAligned(p, v);
+            p += 8;
+        }
+
+        [Benchmark]
+        public void Sse41_V6_Aligned()
+        {
+            float* p = this.alignedPtr;
+
+            Round8SseVectors(p);
+            Round8SseVectors(p + 32);
+        }
+
+        private static void Round8SseVectors(float* p0)
+        {
+            float* p1 = p0 + 4;
+            float* p2 = p1 + 4;
+            float* p3 = p2 + 4;
+            float* p4 = p3 + 4;
+            float* p5 = p4 + 4;
+            float* p6 = p5 + 4;
+            float* p7 = p6 + 4;
+
+            Vector128<float> v0 = Sse.LoadAlignedVector128(p0);
+            Vector128<float> v1 = Sse.LoadAlignedVector128(p1);
+            Vector128<float> v2 = Sse.LoadAlignedVector128(p2);
+            Vector128<float> v3 = Sse.LoadAlignedVector128(p3);
+            Vector128<float> v4 = Sse.LoadAlignedVector128(p4);
+            Vector128<float> v5 = Sse.LoadAlignedVector128(p5);
+            Vector128<float> v6 = Sse.LoadAlignedVector128(p6);
+            Vector128<float> v7 = Sse.LoadAlignedVector128(p7);
+
+            v0 = Sse41.RoundToNearestInteger(v0);
+            v1 = Sse41.RoundToNearestInteger(v1);
+            v2 = Sse41.RoundToNearestInteger(v2);
+            v3 = Sse41.RoundToNearestInteger(v3);
+            v4 = Sse41.RoundToNearestInteger(v4);
+            v5 = Sse41.RoundToNearestInteger(v5);
+            v6 = Sse41.RoundToNearestInteger(v6);
+            v7 = Sse41.RoundToNearestInteger(v7);
+
+            Sse.StoreAligned(p0, v0);
+            Sse.StoreAligned(p1, v1);
+            Sse.StoreAligned(p2, v2);
+            Sse.StoreAligned(p3, v3);
+            Sse.StoreAligned(p4, v4);
+            Sse.StoreAligned(p5, v5);
+            Sse.StoreAligned(p6, v6);
+            Sse.StoreAligned(p7, v7);
+        }
+#endif
     }
 }

tests/ImageSharp.Benchmarks/Codecs/Jpeg/DecodeJpeg_ImageSpecific.cs

@@ -44,6 +44,7 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg
 
         private string TestImageFullPath => Path.Combine(TestEnvironment.InputImagesDirectoryFullPath, this.TestImage);
 
+        #pragma warning disable SA1115
         [Params(
             TestImages.Jpeg.BenchmarkSuite.Lake_Small444YCbCr,
             TestImages.Jpeg.BenchmarkSuite.BadRstProgressive518_Large444YCbCr,

tests/ImageSharp.Benchmarks/Codecs/Jpeg/YCbCrColorConversion.cs

@@ -36,7 +36,7 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg
             }
         }
 
-        [Benchmark(Baseline = true)]
+        [Benchmark]
         public void Scalar()
         {
             var values = new JpegColorConverter.ComponentValues(this.input, 0);
@@ -44,7 +44,7 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg
             JpegColorConverter.FromYCbCrBasic.ConvertCore(values, this.output, 255F, 128F);
         }
 
-        [Benchmark]
+        [Benchmark(Baseline = true)]
         public void SimdVector4()
         {
             var values = new JpegColorConverter.ComponentValues(this.input, 0);
@@ -53,11 +53,11 @@ namespace SixLabors.ImageSharp.Benchmarks.Codecs.Jpeg
         }
 
         [Benchmark]
-        public void SimdAvx2()
+        public void SimdVector8()
         {
             var values = new JpegColorConverter.ComponentValues(this.input, 0);
 
-            JpegColorConverter.FromYCbCrSimdAvx2.ConvertCore(values, this.output, 255F, 128F);
+            JpegColorConverter.FromYCbCrSimdVector8.ConvertCore(values, this.output, 255F, 128F);
         }
 
         private static Buffer2D<float>[] CreateRandomValues(

tests/ImageSharp.Benchmarks/Color/Bulk/FromVector4.cs

@@ -7,8 +7,14 @@ using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 
-using BenchmarkDotNet.Attributes;
+#if SUPPORTS_RUNTIME_INTRINSICS
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+#endif
 
+using BenchmarkDotNet.Attributes;
+using BenchmarkDotNet.Environments;
+using BenchmarkDotNet.Jobs;
 using SixLabors.ImageSharp.Memory;
 using SixLabors.ImageSharp.PixelFormats;
 
@@ -25,7 +31,8 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk
 
         protected Configuration Configuration => Configuration.Default;
 
-        [Params(64, 2048)]
+        // [Params(64, 2048)]
+        [Params(1024)]
         public int Count { get; set; }
 
         [GlobalSetup]
@@ -74,52 +81,105 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk
             Span<float> sBytes = MemoryMarshal.Cast<Vector4, float>(this.source.GetSpan());
             Span<byte> dFloats = MemoryMarshal.Cast<Rgba32, byte>(this.destination.GetSpan());
 
-            SimdUtils.FallbackIntrinsics128.BulkConvertNormalizedFloatToByteClampOverflows(sBytes, dFloats);
+            SimdUtils.FallbackIntrinsics128.NormalizedFloatToByteSaturate(sBytes, dFloats);
         }
 
-        [Benchmark(Baseline = true)]
+        [Benchmark]
         public void BasicIntrinsics256()
         {
             Span<float> sBytes = MemoryMarshal.Cast<Vector4, float>(this.source.GetSpan());
             Span<byte> dFloats = MemoryMarshal.Cast<Rgba32, byte>(this.destination.GetSpan());
 
-            SimdUtils.BasicIntrinsics256.BulkConvertNormalizedFloatToByteClampOverflows(sBytes, dFloats);
+            SimdUtils.BasicIntrinsics256.NormalizedFloatToByteSaturate(sBytes, dFloats);
         }
 
-        [Benchmark]
+        [Benchmark(Baseline = true)]
         public void ExtendedIntrinsic()
         {
             Span<float> sBytes = MemoryMarshal.Cast<Vector4, float>(this.source.GetSpan());
             Span<byte> dFloats = MemoryMarshal.Cast<Rgba32, byte>(this.destination.GetSpan());
 
-            SimdUtils.ExtendedIntrinsics.BulkConvertNormalizedFloatToByteClampOverflows(sBytes, dFloats);
+            SimdUtils.ExtendedIntrinsics.NormalizedFloatToByteSaturate(sBytes, dFloats);
+        }
+
+#if SUPPORTS_RUNTIME_INTRINSICS
+        [Benchmark]
+        public void UseAvx2()
+        {
+            Span<float> sBytes = MemoryMarshal.Cast<Vector4, float>(this.source.GetSpan());
+            Span<byte> dFloats = MemoryMarshal.Cast<Rgba32, byte>(this.destination.GetSpan());
+
+            SimdUtils.Avx2Intrinsics.NormalizedFloatToByteSaturate(sBytes, dFloats);
+        }
+
+        private static ReadOnlySpan<byte> PermuteMaskDeinterleave8x32 => new byte[] { 0, 0, 0, 0, 4, 0, 0, 0, 1, 0, 0, 0, 5, 0, 0, 0, 2, 0, 0, 0, 6, 0, 0, 0, 3, 0, 0, 0, 7, 0, 0, 0 };
+
+        [Benchmark]
+        public void UseAvx2_Grouped()
+        {
+            Span<float> src = MemoryMarshal.Cast<Vector4, float>(this.source.GetSpan());
+            Span<byte> dest = MemoryMarshal.Cast<Rgba32, byte>(this.destination.GetSpan());
+
+            int n = dest.Length / Vector<byte>.Count;
+
+            ref Vector256<float> sourceBase =
+                ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(src));
+            ref Vector256<byte> destBase = ref Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(dest));
+
+            ref byte maskBase = ref MemoryMarshal.GetReference(PermuteMaskDeinterleave8x32);
+            Vector256<int> mask = Unsafe.As<byte, Vector256<int>>(ref maskBase);
+
+            var maxBytes = Vector256.Create(255f);
+
+            for (int i = 0; i < n; i++)
+            {
+                ref Vector256<float> s = ref Unsafe.Add(ref sourceBase, i * 4);
+
+                Vector256<float> f0 = s;
+                Vector256<float> f1 = Unsafe.Add(ref s, 1);
+                Vector256<float> f2 = Unsafe.Add(ref s, 2);
+                Vector256<float> f3 = Unsafe.Add(ref s, 3);
+
+                f0 = Avx.Multiply(maxBytes, f0);
+                f1 = Avx.Multiply(maxBytes, f1);
+                f2 = Avx.Multiply(maxBytes, f2);
+                f3 = Avx.Multiply(maxBytes, f3);
+
+                Vector256<int> w0 = Avx.ConvertToVector256Int32(f0);
+                Vector256<int> w1 = Avx.ConvertToVector256Int32(f1);
+                Vector256<int> w2 = Avx.ConvertToVector256Int32(f2);
+                Vector256<int> w3 = Avx.ConvertToVector256Int32(f3);
+
+                Vector256<short> u0 = Avx2.PackSignedSaturate(w0, w1);
+                Vector256<short> u1 = Avx2.PackSignedSaturate(w2, w3);
+                Vector256<byte> b = Avx2.PackUnsignedSaturate(u0, u1);
+                b = Avx2.PermuteVar8x32(b.AsInt32(), mask).AsByte();
+
+                Unsafe.Add(ref destBase, i) = b;
+            }
         }
 
-        // RESULTS (2018 October):
-        //                       Method | Runtime | Count |         Mean |        Error |      StdDev | Scaled | ScaledSD |  Gen 0 | Allocated |
-        // ---------------------------- |-------- |------ |-------------:|-------------:|------------:|-------:|---------:|-------:|----------:|
-        //        FallbackIntrinsics128 |     Clr |    64 |    340.38 ns |    22.319 ns |   1.2611 ns |   1.41 |     0.01 |      - |       0 B |
-        //           BasicIntrinsics256 |     Clr |    64 |    240.79 ns |    11.421 ns |   0.6453 ns |   1.00 |     0.00 |      - |       0 B |
-        //            ExtendedIntrinsic |     Clr |    64 |    199.09 ns |   124.239 ns |   7.0198 ns |   0.83 |     0.02 |      - |       0 B |
-        //         PixelOperations_Base |     Clr |    64 |    647.99 ns |    24.003 ns |   1.3562 ns |   2.69 |     0.01 | 0.0067 |      24 B |
-        //  PixelOperations_Specialized |     Clr |    64 |    259.79 ns |    13.391 ns |   0.7566 ns |   1.08 |     0.00 |      - |       0 B | <--- ceremonial overhead has been minimized!
-        //                              |         |       |              |              |             |        |          |        |           |
-        //        FallbackIntrinsics128 |    Core |    64 |    234.64 ns |    12.320 ns |   0.6961 ns |   1.58 |     0.00 |      - |       0 B |
-        //           BasicIntrinsics256 |    Core |    64 |    148.87 ns |     2.794 ns |   0.1579 ns |   1.00 |     0.00 |      - |       0 B |
-        //            ExtendedIntrinsic |    Core |    64 |     94.06 ns |    10.015 ns |   0.5659 ns |   0.63 |     0.00 |      - |       0 B |
-        //         PixelOperations_Base |    Core |    64 |    573.52 ns |    31.865 ns |   1.8004 ns |   3.85 |     0.01 | 0.0067 |      24 B |
-        //  PixelOperations_Specialized |    Core |    64 |    117.21 ns |    13.264 ns |   0.7494 ns |   0.79 |     0.00 |      - |       0 B |
-        //                              |         |       |              |              |             |        |          |        |           |
-        //        FallbackIntrinsics128 |     Clr |  2048 |  6,735.93 ns | 2,139.340 ns | 120.8767 ns |   1.71 |     0.03 |      - |       0 B |
-        //           BasicIntrinsics256 |     Clr |  2048 |  3,929.29 ns |   334.027 ns |  18.8731 ns |   1.00 |     0.00 |      - |       0 B |
-        //            ExtendedIntrinsic |     Clr |  2048 |  2,226.01 ns |   130.525 ns |   7.3749 ns |!! 0.57 |     0.00 |      - |       0 B | <--- ExtendedIntrinsics rock!
-        //         PixelOperations_Base |     Clr |  2048 | 16,760.84 ns |   367.800 ns |  20.7814 ns |   4.27 |     0.02 |      - |      24 B | <--- Extra copies using "Vector4 TPixel.ToVector4()"
-        //  PixelOperations_Specialized |     Clr |  2048 |  3,986.03 ns |   237.238 ns |  13.4044 ns |   1.01 |     0.00 |      - |       0 B | <--- can't yet detect whether ExtendedIntrinsics are available :(
-        //                              |         |       |              |              |             |        |          |        |           |
-        //        FallbackIntrinsics128 |    Core |  2048 |  6,644.65 ns | 2,677.090 ns | 151.2605 ns |   1.69 |     0.05 |      - |       0 B |
-        //           BasicIntrinsics256 |    Core |  2048 |  3,923.70 ns | 1,971.760 ns | 111.4081 ns |   1.00 |     0.00 |      - |       0 B |
-        //            ExtendedIntrinsic |    Core |  2048 |  2,092.32 ns |   375.657 ns |  21.2253 ns |!! 0.53 |     0.01 |      - |       0 B | <--- ExtendedIntrinsics rock!
-        //         PixelOperations_Base |    Core |  2048 | 16,875.73 ns | 1,271.957 ns |  71.8679 ns |   4.30 |     0.10 |      - |      24 B |
-        //  PixelOperations_Specialized |    Core |  2048 |  2,129.92 ns |   262.888 ns |  14.8537 ns |!! 0.54 |     0.01 |      - |       0 B | <--- ExtendedIntrinsics rock!
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private static Vector256<int> ConvertToInt32(Vector256<float> vf, Vector256<float> scale)
+        {
+            vf = Avx.Multiply(scale, vf);
+            return Avx.ConvertToVector256Int32(vf);
+        }
+#endif
+
+        // *** RESULTS 2020 March: ***
+        // Intel Core i7-8650U CPU 1.90GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores
+        // .NET Core SDK=3.1.200-preview-014971
+        //   Job-IUZXZT : .NET Core 3.1.2 (CoreCLR 4.700.20.6602, CoreFX 4.700.20.6702), X64 RyuJIT
+        //
+        // |                      Method | Count |       Mean |       Error |    StdDev | Ratio | RatioSD | Gen 0 | Gen 1 | Gen 2 | Allocated |
+        // |---------------------------- |------ |-----------:|------------:|----------:|------:|--------:|------:|------:|------:|----------:|
+        // |       FallbackIntrinsics128 |  1024 | 2,952.6 ns | 1,680.77 ns |  92.13 ns |  3.32 |    0.16 |     - |     - |     - |         - |
+        // |          BasicIntrinsics256 |  1024 | 1,664.5 ns |   928.11 ns |  50.87 ns |  1.87 |    0.09 |     - |     - |     - |         - |
+        // |           ExtendedIntrinsic |  1024 |   890.6 ns |   375.48 ns |  20.58 ns |  1.00 |    0.00 |     - |     - |     - |         - |
+        // |                     UseAvx2 |  1024 |   299.0 ns |    30.47 ns |   1.67 ns |  0.34 |    0.01 |     - |     - |     - |         - |
+        // |             UseAvx2_Grouped |  1024 |   318.1 ns |    48.19 ns |   2.64 ns |  0.36 |    0.01 |     - |     - |     - |         - |
+        // |        PixelOperations_Base |  1024 | 8,136.9 ns | 1,834.82 ns | 100.57 ns |  9.14 |    0.26 |     - |     - |     - |      24 B |
+        // | PixelOperations_Specialized |  1024 |   951.1 ns |   123.93 ns |   6.79 ns |  1.07 |    0.03 |     - |     - |     - |         - |
     }
 }

tests/ImageSharp.Benchmarks/Color/Bulk/ToVector4_Rgba32.cs

@@ -22,7 +22,7 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk
             Span<byte> sBytes = MemoryMarshal.Cast<Rgba32, byte>(this.source.GetSpan());
             Span<float> dFloats = MemoryMarshal.Cast<Vector4, float>(this.destination.GetSpan());
 
-            SimdUtils.FallbackIntrinsics128.BulkConvertByteToNormalizedFloat(sBytes, dFloats);
+            SimdUtils.FallbackIntrinsics128.ByteToNormalizedFloat(sBytes, dFloats);
         }
 
         [Benchmark]
@@ -40,7 +40,7 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk
             Span<byte> sBytes = MemoryMarshal.Cast<Rgba32, byte>(this.source.GetSpan());
             Span<float> dFloats = MemoryMarshal.Cast<Vector4, float>(this.destination.GetSpan());
 
-            SimdUtils.BasicIntrinsics256.BulkConvertByteToNormalizedFloat(sBytes, dFloats);
+            SimdUtils.BasicIntrinsics256.ByteToNormalizedFloat(sBytes, dFloats);
         }
 
         [Benchmark]
@@ -49,7 +49,7 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk
             Span<byte> sBytes = MemoryMarshal.Cast<Rgba32, byte>(this.source.GetSpan());
             Span<float> dFloats = MemoryMarshal.Cast<Vector4, float>(this.destination.GetSpan());
 
-            SimdUtils.ExtendedIntrinsics.BulkConvertByteToNormalizedFloat(sBytes, dFloats);
+            SimdUtils.ExtendedIntrinsics.ByteToNormalizedFloat(sBytes, dFloats);
         }
 
         // [Benchmark]

tests/ImageSharp.Benchmarks/Config.cs

@@ -38,6 +38,14 @@ namespace SixLabors.ImageSharp.Benchmarks
             }
         }
 
+        public class ShortCore31 : Config
+        {
+            public ShortCore31()
+            {
+                this.Add(Job.Default.With(CoreRuntime.Core31).WithLaunchCount(1).WithWarmupCount(3).WithIterationCount(3));
+            }
+        }
+
 #if Windows_NT
         private bool IsElevated
         {

tests/ImageSharp.Tests.ProfilingSandbox/Program.cs

@@ -7,6 +7,10 @@ using SixLabors.ImageSharp.Tests.PixelFormats.PixelOperations;
 using SixLabors.ImageSharp.Tests.ProfilingBenchmarks;
 using Xunit.Abstractions;
 
+// in this file, comments are used for disabling stuff for local execution
+#pragma warning disable SA1515
+#pragma warning disable SA1512
+
 namespace SixLabors.ImageSharp.Tests.ProfilingSandbox
 {
     public class Program
@@ -28,10 +32,9 @@ namespace SixLabors.ImageSharp.Tests.ProfilingSandbox
         public static void Main(string[] args)
         {
             // RunJpegColorProfilingTests();
-
-            // RunDecodeJpegProfilingTests();
+            RunDecodeJpegProfilingTests();
             // RunToVector4ProfilingTest();
-            RunResizeProfilingTest();
+            // RunResizeProfilingTest();
 
             Console.ReadLine();
         }
@@ -61,8 +64,11 @@ namespace SixLabors.ImageSharp.Tests.ProfilingSandbox
             foreach (object[] data in JpegProfilingBenchmarks.DecodeJpegData)
             {
                 string fileName = (string)data[0];
-                benchmarks.DecodeJpeg(fileName);
+                int executionCount = (int)data[1];
+                benchmarks.DecodeJpeg(fileName, executionCount);
             }
+
+            Console.WriteLine("DONE.");
         }
     }
 }

tests/ImageSharp.Tests/Common/SimdUtilsTests.cs

@@ -105,7 +105,7 @@ namespace SixLabors.ImageSharp.Tests.Common
 
         private bool SkipOnNonAvx2([CallerMemberName] string testCaseName = null)
         {
-            if (!SimdUtils.IsAvx2CompatibleArchitecture)
+            if (!SimdUtils.HasVector8)
             {
                 this.Output.WriteLine("Skipping AVX2 specific test case: " + testCaseName);
                 return true;
@@ -178,7 +178,7 @@ namespace SixLabors.ImageSharp.Tests.Common
         {
             TestImpl_BulkConvertByteToNormalizedFloat(
                 count,
-                (s, d) => SimdUtils.FallbackIntrinsics128.BulkConvertByteToNormalizedFloat(s.Span, d.Span));
+                (s, d) => SimdUtils.FallbackIntrinsics128.ByteToNormalizedFloat(s.Span, d.Span));
         }
 
         [Theory]
@@ -192,7 +192,7 @@ namespace SixLabors.ImageSharp.Tests.Common
 
             TestImpl_BulkConvertByteToNormalizedFloat(
                 count,
-                (s, d) => SimdUtils.BasicIntrinsics256.BulkConvertByteToNormalizedFloat(s.Span, d.Span));
+                (s, d) => SimdUtils.BasicIntrinsics256.ByteToNormalizedFloat(s.Span, d.Span));
         }
 
         [Theory]
@@ -201,7 +201,7 @@ namespace SixLabors.ImageSharp.Tests.Common
         {
             TestImpl_BulkConvertByteToNormalizedFloat(
                 count,
-                (s, d) => SimdUtils.ExtendedIntrinsics.BulkConvertByteToNormalizedFloat(s.Span, d.Span));
+                (s, d) => SimdUtils.ExtendedIntrinsics.ByteToNormalizedFloat(s.Span, d.Span));
         }
 
         [Theory]
@@ -210,7 +210,7 @@ namespace SixLabors.ImageSharp.Tests.Common
         {
             TestImpl_BulkConvertByteToNormalizedFloat(
                 count,
-                (s, d) => SimdUtils.BulkConvertByteToNormalizedFloat(s.Span, d.Span));
+                (s, d) => SimdUtils.ByteToNormalizedFloat(s.Span, d.Span));
         }
 
         private static void TestImpl_BulkConvertByteToNormalizedFloat(
@@ -232,7 +232,7 @@ namespace SixLabors.ImageSharp.Tests.Common
         {
             TestImpl_BulkConvertNormalizedFloatToByteClampOverflows(
                 count,
-                (s, d) => SimdUtils.FallbackIntrinsics128.BulkConvertNormalizedFloatToByteClampOverflows(s.Span, d.Span));
+                (s, d) => SimdUtils.FallbackIntrinsics128.NormalizedFloatToByteSaturate(s.Span, d.Span));
         }
 
         [Theory]
@@ -244,7 +244,7 @@ namespace SixLabors.ImageSharp.Tests.Common
                 return;
             }
 
-            TestImpl_BulkConvertNormalizedFloatToByteClampOverflows(count, (s, d) => SimdUtils.BasicIntrinsics256.BulkConvertNormalizedFloatToByteClampOverflows(s.Span, d.Span));
+            TestImpl_BulkConvertNormalizedFloatToByteClampOverflows(count, (s, d) => SimdUtils.BasicIntrinsics256.NormalizedFloatToByteSaturate(s.Span, d.Span));
         }
 
         [Theory]
@@ -253,7 +253,7 @@ namespace SixLabors.ImageSharp.Tests.Common
         {
             TestImpl_BulkConvertNormalizedFloatToByteClampOverflows(
                 count,
-                (s, d) => SimdUtils.ExtendedIntrinsics.BulkConvertNormalizedFloatToByteClampOverflows(s.Span, d.Span));
+                (s, d) => SimdUtils.ExtendedIntrinsics.NormalizedFloatToByteSaturate(s.Span, d.Span));
         }
 
         [Theory]
@@ -277,11 +277,29 @@ namespace SixLabors.ImageSharp.Tests.Common
             Assert.Equal(expected2, actual2);
         }
 
+#if SUPPORTS_RUNTIME_INTRINSICS
+
+        [Theory]
+        [MemberData(nameof(ArraySizesDivisibleBy32))]
+        public void Avx2_BulkConvertNormalizedFloatToByteClampOverflows(int count)
+        {
+            if (!System.Runtime.Intrinsics.X86.Avx2.IsSupported)
+            {
+                return;
+            }
+
+            TestImpl_BulkConvertNormalizedFloatToByteClampOverflows(
+                count,
+                (s, d) => SimdUtils.Avx2Intrinsics.NormalizedFloatToByteSaturate(s.Span, d.Span));
+        }
+
+#endif
+
         [Theory]
         [MemberData(nameof(ArbitraryArraySizes))]
         public void BulkConvertNormalizedFloatToByteClampOverflows(int count)
         {
-            TestImpl_BulkConvertNormalizedFloatToByteClampOverflows(count, (s, d) => SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows(s.Span, d.Span));
+            TestImpl_BulkConvertNormalizedFloatToByteClampOverflows(count, (s, d) => SimdUtils.NormalizedFloatToByteSaturate(s.Span, d.Span));
 
             // For small values, let's stress test the implementation a bit:
             if (count > 0 && count < 10)
@@ -290,7 +308,7 @@ namespace SixLabors.ImageSharp.Tests.Common
                 {
                     TestImpl_BulkConvertNormalizedFloatToByteClampOverflows(
                         count,
-                        (s, d) => SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows(s.Span, d.Span),
+                        (s, d) => SimdUtils.NormalizedFloatToByteSaturate(s.Span, d.Span),
                         i + 42);
                 }
             }

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

@@ -44,7 +44,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 using (Buffer2D<float> buffer = Configuration.Default.MemoryAllocator.Allocate2D<float>(20, 20, AllocationOptions.Clean))
                 {
                     BufferArea<float> area = buffer.GetArea(5, 10, 8, 8);
-                    block.Copy1x1Scale(area);
+                    block.Copy1x1Scale(ref area.GetReferenceToOrigin(), area.Stride);
 
                     Assert.Equal(block[0, 0], buffer[5, 10]);
                     Assert.Equal(block[1, 0], buffer[6, 10]);
@@ -72,7 +72,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 using (Buffer2D<float> buffer = Configuration.Default.MemoryAllocator.Allocate2D<float>(100, 100, AllocationOptions.Clean))
                 {
                     BufferArea<float> area = buffer.GetArea(start.X, start.Y, 8 * horizontalFactor, 8 * verticalFactor);
-                    block.CopyTo(area, horizontalFactor, verticalFactor);
+                    block.ScaledCopyTo(area, horizontalFactor, verticalFactor);
 
                     for (int y = 0; y < 8 * verticalFactor; y++)
                     {

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

@@ -29,7 +29,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
         private bool SkipOnNonAvx2Runner()
         {
-            if (!SimdUtils.IsAvx2CompatibleArchitecture)
+            if (!SimdUtils.HasVector8)
             {
                 this.Output.WriteLine("AVX2 not supported, skipping!");
                 return true;
@@ -104,7 +104,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                     {
                         var b = default(Block8x8F);
                         b.LoadFrom(data);
-                        b.CopyTo(mirror);
+                        b.ScaledCopyTo(mirror);
                     });
 
             Assert.Equal(data, mirror);
@@ -129,7 +129,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                     {
                         var b = default(Block8x8F);
                         Block8x8F.LoadFrom(&b, data);
-                        Block8x8F.CopyTo(&b, mirror);
+                        Block8x8F.ScaledCopyTo(&b, mirror);
                     });
 
             Assert.Equal(data, mirror);
@@ -154,7 +154,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                     {
                         var v = default(Block8x8F);
                         v.LoadFrom(data);
-                        v.CopyTo(mirror);
+                        v.ScaledCopyTo(mirror);
                     });
 
             Assert.Equal(data, mirror);
@@ -175,7 +175,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             source.TransposeInto(ref dest);
 
             float[] actual = new float[64];
-            dest.CopyTo(actual);
+            dest.ScaledCopyTo(actual);
 
             Assert.Equal(expected, actual);
         }
@@ -231,7 +231,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             dest.NormalizeColorsInplace(255);
 
             float[] array = new float[64];
-            dest.CopyTo(array);
+            dest.ScaledCopyTo(array);
             this.Output.WriteLine("Result:");
             this.PrintLinearData(array);
             foreach (float val in array)
@@ -257,7 +257,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
             expected.RoundInplace();
 
             Block8x8F actual = source;
-            actual.NormalizeColorsAndRoundInplaceAvx2(255);
+            actual.NormalizeColorsAndRoundInplaceVector8(255);
 
             this.Output.WriteLine(expected.ToString());
             this.Output.WriteLine(actual.ToString());

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

@@ -37,7 +37,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 FastFloatingPointDCT.IDCT8x4_LeftPart(ref source, ref dest);
 
                 var actualDestArray = new float[64];
-                dest.CopyTo(actualDestArray);
+                dest.ScaledCopyTo(actualDestArray);
 
                 this.Print8x8Data(expectedDestArray);
                 this.Output.WriteLine("**************");
@@ -62,7 +62,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 FastFloatingPointDCT.IDCT8x4_RightPart(ref source, ref dest);
 
                 var actualDestArray = new float[64];
-                dest.CopyTo(actualDestArray);
+                dest.ScaledCopyTo(actualDestArray);
 
                 this.Print8x8Data(expectedDestArray);
                 this.Output.WriteLine("**************");
@@ -126,7 +126,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 FastFloatingPointDCT.FDCT8x4_LeftPart(ref srcBlock, ref destBlock);
 
                 var actualDest = new float[64];
-                destBlock.CopyTo(actualDest);
+                destBlock.ScaledCopyTo(actualDest);
 
                 Assert.Equal(actualDest, expectedDest, new ApproximateFloatComparer(1f));
             }
@@ -148,7 +148,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 FastFloatingPointDCT.FDCT8x4_RightPart(ref srcBlock, ref destBlock);
 
                 var actualDest = new float[64];
-                destBlock.CopyTo(actualDest);
+                destBlock.ScaledCopyTo(actualDest);
 
                 Assert.Equal(actualDest, expectedDest, new ApproximateFloatComparer(1f));
             }
@@ -172,7 +172,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
                 FastFloatingPointDCT.TransformFDCT(ref srcBlock, ref destBlock, ref temp2, false);
 
                 var actualDest = new float[64];
-                destBlock.CopyTo(actualDest);
+                destBlock.ScaledCopyTo(actualDest);
 
                 Assert.Equal(actualDest, expectedDest, new ApproximateFloatComparer(1f));
             }

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

@@ -99,7 +99,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
         [MemberData(nameof(CommonConversionData))]
         public void FromYCbCrSimdAvx2(int inputBufferLength, int resultBufferLength, int seed)
         {
-            if (!SimdUtils.IsAvx2CompatibleArchitecture)
+            if (!SimdUtils.HasVector8)
             {
                 this.Output.WriteLine("No AVX2 present, skipping test!");
                 return;
@@ -107,7 +107,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg
 
             // JpegColorConverter.FromYCbCrSimdAvx2.LogPlz = s => this.Output.WriteLine(s);
             ValidateRgbToYCbCrConversion(
-                new JpegColorConverter.FromYCbCrSimdAvx2(8),
+                new JpegColorConverter.FromYCbCrSimdVector8(8),
                 3,
                 inputBufferLength,
                 resultBufferLength,

tests/ImageSharp.Tests/Formats/Jpg/Utils/ReferenceImplementations.LLM_FloatingPoint_DCT.cs

@@ -33,7 +33,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg.Utils
             public static Block8x8F TransformIDCT(ref Block8x8F source)
             {
                 float[] s = new float[64];
-                source.CopyTo(s);
+                source.ScaledCopyTo(s);
                 float[] d = new float[64];
                 float[] temp = new float[64];
 
@@ -46,7 +46,7 @@ namespace SixLabors.ImageSharp.Tests.Formats.Jpg.Utils
             public static Block8x8F TransformFDCT_UpscaleBy8(ref Block8x8F source)
             {
                 float[] s = new float[64];
-                source.CopyTo(s);
+                source.ScaledCopyTo(s);
                 float[] d = new float[64];
                 float[] temp = new float[64];
 

tests/ImageSharp.Tests/Processing/Processors/Convolution/DetectEdgesTest.cs

@@ -12,9 +12,9 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Convolution
     [GroupOutput("Convolution")]
     public class DetectEdgesTest
     {
-        // I think our comparison is not accurate enough (nor can be) for RgbaVector.
-        // The image pixels are identical according to BeyondCompare.
-        private static readonly ImageComparer ValidatorComparer = ImageComparer.TolerantPercentage(0.0456F);
+        private static readonly ImageComparer OpaqueComparer = ImageComparer.TolerantPercentage(0.01F);
+
+        private static readonly ImageComparer TransparentComparer = ImageComparer.TolerantPercentage(0.5F);
 
         public static readonly string[] TestImages = { Tests.TestImages.Png.Bike };
 
@@ -46,7 +46,7 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Convolution
                         var bounds = new Rectangle(10, 10, size.Width / 2, size.Height / 2);
                         ctx.DetectEdges(bounds);
                     },
-                comparer: ValidatorComparer,
+                comparer: OpaqueComparer,
                 useReferenceOutputFrom: nameof(this.DetectEdges_InBox));
         }
 
@@ -56,11 +56,13 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Convolution
         public void DetectEdges_WorksWithAllFilters<TPixel>(TestImageProvider<TPixel> provider, EdgeDetectionOperators detector)
             where TPixel : unmanaged, IPixel<TPixel>
         {
+            bool hasAlpha = provider.SourceFileOrDescription.Contains("TestPattern");
+            ImageComparer comparer = hasAlpha ? TransparentComparer : OpaqueComparer;
             using (Image<TPixel> image = provider.GetImage())
             {
                 image.Mutate(x => x.DetectEdges(detector));
                 image.DebugSave(provider, detector.ToString());
-                image.CompareToReferenceOutput(ValidatorComparer, provider, detector.ToString());
+                image.CompareToReferenceOutput(comparer, provider, detector.ToString());
             }
         }
 
@@ -69,11 +71,18 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Convolution
         public void DetectEdges_IsNotBoundToSinglePixelType<TPixel>(TestImageProvider<TPixel> provider)
             where TPixel : unmanaged, IPixel<TPixel>
         {
+            // James:
+            // I think our comparison is not accurate enough (nor can be) for RgbaVector.
+            // The image pixels are identical according to BeyondCompare.
+            ImageComparer comparer = typeof(TPixel) == typeof(RgbaVector) ?
+                ImageComparer.TolerantPercentage(1f) :
+                OpaqueComparer;
+
             using (Image<TPixel> image = provider.GetImage())
             {
                 image.Mutate(x => x.DetectEdges());
                 image.DebugSave(provider);
-                image.CompareToReferenceOutput(ValidatorComparer, provider);
+                image.CompareToReferenceOutput(comparer, provider);
             }
         }
 
@@ -100,7 +109,7 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Convolution
 
                 image.Mutate(x => x.DetectEdges(bounds));
                 image.DebugSave(provider);
-                image.CompareToReferenceOutput(ValidatorComparer, provider);
+                image.CompareToReferenceOutput(OpaqueComparer, provider);
             }
         }
 

tests/ImageSharp.Tests/Processing/Processors/Effects/OilPaintTest.cs

@@ -3,7 +3,7 @@
 
 using SixLabors.ImageSharp.PixelFormats;
 using SixLabors.ImageSharp.Processing;
-
+using SixLabors.ImageSharp.Tests.TestUtilities.ImageComparison;
 using Xunit;
 
 namespace SixLabors.ImageSharp.Tests.Processing.Processors.Effects
@@ -29,8 +29,12 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Effects
             where TPixel : unmanaged, IPixel<TPixel>
         {
             provider.RunValidatingProcessorTest(
-                x => x.OilPaint(levels, brushSize),
-                $"{levels}-{brushSize}",
+                x =>
+                {
+                    x.OilPaint(levels, brushSize);
+                    return $"{levels}-{brushSize}";
+                },
+                ImageComparer.TolerantPercentage(0.01F),
                 appendPixelTypeToFileName: false);
         }
 
@@ -42,7 +46,8 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Effects
         {
             provider.RunRectangleConstrainedValidatingProcessorTest(
                 (x, rect) => x.OilPaint(levels, brushSize, rect),
-                $"{levels}-{brushSize}");
+                $"{levels}-{brushSize}",
+                ImageComparer.TolerantPercentage(0.01F));
         }
     }
 }

tests/ImageSharp.Tests/Processing/Processors/Quantization/QuantizerTests.cs

@@ -19,7 +19,7 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Quantization
         /// Not worth investigating for now.
         /// <see href="https://github.com/SixLabors/ImageSharp/pull/1114/checks?check_run_id=448891164#step:11:631"/>
         /// </summary>
-        private static readonly bool SkipAllQuantizerTests = TestEnvironment.RunsOnCI && TestEnvironment.IsFramework;
+        private static readonly bool SkipAllQuantizerTests = TestEnvironment.IsFramework;
 
         public static readonly string[] CommonTestImages =
         {

tests/ImageSharp.Tests/Processing/Processors/Transforms/EntropyCropTest.cs

@@ -1,6 +1,7 @@
 // Copyright (c) Six Labors and contributors.
 // Licensed under the Apache License, Version 2.0.
 
+using System;
 using SixLabors.ImageSharp.PixelFormats;
 using SixLabors.ImageSharp.Processing;
 using Xunit;
@@ -24,7 +25,16 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Transforms
         public void EntropyCrop<TPixel>(TestImageProvider<TPixel> provider, float value)
             where TPixel : unmanaged, IPixel<TPixel>
         {
+            // The result dimensions of EntropyCrop may differ on .NET Core 3.1 because of unstable edge detection results.
+            // TODO: Re-enable this test case if we manage to improve stability.
+#if SUPPORTS_RUNTIME_INTRINSICS
+            if (provider.SourceFileOrDescription.Contains(TestImages.Png.Ducky))
+            {
+                return;
+            }
+#endif
+
             provider.RunValidatingProcessorTest(x => x.EntropyCrop(value), value, appendPixelTypeToFileName: false);
         }
     }
-}
+}

tests/ImageSharp.Tests/ProfilingBenchmarks/JpegProfilingBenchmarks.cs

@@ -13,6 +13,9 @@ using SixLabors.ImageSharp.PixelFormats;
 using Xunit;
 using Xunit.Abstractions;
 
+// in this file, comments are used for disabling stuff for local execution
+#pragma warning disable SA1515
+
 namespace SixLabors.ImageSharp.Tests.ProfilingBenchmarks
 {
     public class JpegProfilingBenchmarks : MeasureFixture
@@ -22,24 +25,28 @@ namespace SixLabors.ImageSharp.Tests.ProfilingBenchmarks
         {
         }
 
-        public static readonly TheoryData<string> DecodeJpegData = new TheoryData<string>
+        public static readonly TheoryData<string, int> DecodeJpegData = new TheoryData<string, int>
         {
-            TestImages.Jpeg.BenchmarkSuite.Jpeg400_SmallMonochrome,
-            TestImages.Jpeg.BenchmarkSuite.Jpeg420Exif_MidSizeYCbCr,
-            TestImages.Jpeg.BenchmarkSuite.Lake_Small444YCbCr,
-            TestImages.Jpeg.BenchmarkSuite.MissingFF00ProgressiveBedroom159_MidSize420YCbCr,
-            TestImages.Jpeg.BenchmarkSuite.BadRstProgressive518_Large444YCbCr,
-            TestImages.Jpeg.BenchmarkSuite.ExifGetString750Transform_Huge420YCbCr,
+            { TestImages.Jpeg.BenchmarkSuite.Jpeg400_SmallMonochrome, 20 },
+            { TestImages.Jpeg.BenchmarkSuite.Jpeg420Exif_MidSizeYCbCr, 20 },
+            { TestImages.Jpeg.BenchmarkSuite.Lake_Small444YCbCr, 40 },
+            // { TestImages.Jpeg.BenchmarkSuite.MissingFF00ProgressiveBedroom159_MidSize420YCbCr, 10 },
+            // { TestImages.Jpeg.BenchmarkSuite.BadRstProgressive518_Large444YCbCr, 5 },
+            { TestImages.Jpeg.BenchmarkSuite.ExifGetString750Transform_Huge420YCbCr, 5 }
         };
 
         [Theory(Skip = ProfilingSetup.SkipProfilingTests)]
         [MemberData(nameof(DecodeJpegData))]
-        public void DecodeJpeg(string fileName)
+        public void DecodeJpeg(string fileName, int executionCount)
         {
-            this.DecodeJpegBenchmarkImpl(fileName, new JpegDecoder());
+            var decoder = new JpegDecoder()
+            {
+                IgnoreMetadata = true
+            };
+            this.DecodeJpegBenchmarkImpl(fileName, decoder, executionCount);
         }
 
-        private void DecodeJpegBenchmarkImpl(string fileName, IImageDecoder decoder)
+        private void DecodeJpegBenchmarkImpl(string fileName, IImageDecoder decoder, int executionCount)
         {
             // do not run this on CI even by accident
             if (TestEnvironment.RunsOnCI)
@@ -47,8 +54,6 @@ namespace SixLabors.ImageSharp.Tests.ProfilingBenchmarks
                 return;
             }
 
-            const int ExecutionCount = 20;
-
             if (!Vector.IsHardwareAccelerated)
             {
                 throw new Exception("Vector.IsHardwareAccelerated == false! ('prefer32 bit' enabled?)");
@@ -58,7 +63,7 @@ namespace SixLabors.ImageSharp.Tests.ProfilingBenchmarks
             byte[] bytes = File.ReadAllBytes(path);
 
             this.Measure(
-                ExecutionCount,
+                executionCount,
                 () =>
                     {
                         var img = Image.Load<Rgba32>(bytes, decoder);