From a4ad7b09fc8dda2f0efa9e0f7abcf4e6f0a4ac67 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?G=C3=BCnther=20Foidl?= <gue@korporal.at>
Date: Sun, 26 Mar 2023 14:20:24 +0200
Subject: [PATCH 1/6] Avoid length-check in pinning spans

Only where it seems profitable. E.g. not when a UnmanagedMemoryStream is constructed of that pointer.
---
 .../ColorSpaces/Companding/SRgbCompanding.cs           |  4 ++--
 .../Common/Helpers/SimdUtils.HwIntrinsics.cs           |  2 +-
 src/ImageSharp/Compression/Zlib/Adler32.cs             |  2 +-
 src/ImageSharp/Compression/Zlib/Crc32.cs               |  6 +++---
 src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs  |  9 ++++-----
 .../Formats/Webp/Lossless/PredictorEncoder.cs          | 10 +++++-----
 src/ImageSharp/Formats/Webp/WebpCommonUtils.cs         |  4 ++--
 src/ImageSharp/Memory/Buffer2DExtensions.cs            |  2 +-
 8 files changed, 19 insertions(+), 20 deletions(-)

diff --git a/src/ImageSharp/ColorSpaces/Companding/SRgbCompanding.cs b/src/ImageSharp/ColorSpaces/Companding/SRgbCompanding.cs
index 2a6fcb0832..4c3923c888 100644
--- a/src/ImageSharp/ColorSpaces/Companding/SRgbCompanding.cs
+++ b/src/ImageSharp/ColorSpaces/Companding/SRgbCompanding.cs
@@ -167,7 +167,7 @@ public static float Compress(float channel)
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     private static unsafe void CompandAvx2(Span<Vector4> vectors, float[] table)
     {
-        fixed (float* tablePointer = &table[0])
+        fixed (float* tablePointer = &MemoryMarshal.GetArrayDataReference(table))
         {
             var scale = Vector256.Create((float)Scale);
             Vector256<float> zero = Vector256<float>.Zero;
@@ -199,7 +199,7 @@ private static unsafe void CompandAvx2(Span<Vector4> vectors, float[] table)
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     private static unsafe void CompandScalar(Span<Vector4> vectors, float[] table)
     {
-        fixed (float* tablePointer = &table[0])
+        fixed (float* tablePointer = &MemoryMarshal.GetArrayDataReference(table))
         {
             Vector4 zero = Vector4.Zero;
             var scale = new Vector4(Scale);
diff --git a/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs b/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs
index ce6f335a8f..7871eee6ba 100644
--- a/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs
+++ b/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs
@@ -644,7 +644,7 @@ internal static unsafe void ByteToNormalizedFloat(
             ReadOnlySpan<byte> source,
             Span<float> dest)
         {
-            fixed (byte* sourceBase = source)
+            fixed (byte* sourceBase = &MemoryMarshal.GetReference(source))
             {
                 if (Avx2.IsSupported)
                 {
diff --git a/src/ImageSharp/Compression/Zlib/Adler32.cs b/src/ImageSharp/Compression/Zlib/Adler32.cs
index dd8217541a..3ecdf81539 100644
--- a/src/ImageSharp/Compression/Zlib/Adler32.cs
+++ b/src/ImageSharp/Compression/Zlib/Adler32.cs
@@ -387,7 +387,7 @@ private static unsafe uint CalculateScalar(uint adler, ReadOnlySpan<byte> buffer
         uint s1 = adler & 0xFFFF;
         uint s2 = (adler >> 16) & 0xFFFF;
 
-        fixed (byte* bufferPtr = buffer)
+        fixed (byte* bufferPtr = &MemoryMarshal.GetReference(buffer))
         {
             byte* localBufferPtr = bufferPtr;
             uint length = (uint)buffer.Length;
diff --git a/src/ImageSharp/Compression/Zlib/Crc32.cs b/src/ImageSharp/Compression/Zlib/Crc32.cs
index 2d0a09bd4c..e21621ab74 100644
--- a/src/ImageSharp/Compression/Zlib/Crc32.cs
+++ b/src/ImageSharp/Compression/Zlib/Crc32.cs
@@ -81,7 +81,7 @@ private static unsafe uint CalculateSse(uint crc, ReadOnlySpan<byte> buffer)
         int chunksize = buffer.Length & ~ChunksizeMask;
         int length = chunksize;
 
-        fixed (byte* bufferPtr = buffer)
+        fixed (byte* bufferPtr = &MemoryMarshal.GetReference(buffer))
         {
             fixed (ulong* k05PolyPtr = K05Poly)
             {
@@ -201,7 +201,7 @@ private static unsafe uint CalculateSse(uint crc, ReadOnlySpan<byte> buffer)
     [MethodImpl(InliningOptions.HotPath | InliningOptions.ShortMethod)]
     private static unsafe uint CalculateArm(uint crc, ReadOnlySpan<byte> buffer)
     {
-        fixed (byte* bufferPtr = buffer)
+        fixed (byte* bufferPtr = &MemoryMarshal.GetReference(buffer))
         {
             byte* localBufferPtr = bufferPtr;
             int len = buffer.Length;
@@ -248,7 +248,7 @@ private static unsafe uint CalculateArm(uint crc, ReadOnlySpan<byte> buffer)
     [MethodImpl(InliningOptions.HotPath | InliningOptions.ShortMethod)]
     private static unsafe uint CalculateArm64(uint crc, ReadOnlySpan<byte> buffer)
     {
-        fixed (byte* bufferPtr = buffer)
+        fixed (byte* bufferPtr = &MemoryMarshal.GetReference(buffer))
         {
             byte* localBufferPtr = bufferPtr;
             int len = buffer.Length;
diff --git a/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs b/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
index 8d64a09dfb..ab6e4cfccf 100644
--- a/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
+++ b/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
@@ -569,7 +569,7 @@ public static void PredictorInverseTransform(
         Span<uint> pixelData,
         Span<uint> outputSpan)
     {
-        fixed (uint* inputFixed = pixelData)
+        fixed (uint* inputFixed = &MemoryMarshal.GetReference(pixelData))
         {
             fixed (uint* outputFixed = outputSpan)
             {
@@ -1474,8 +1474,7 @@ private static uint Select(uint a, uint b, uint c, Span<short> scratch)
     {
         if (Sse2.IsSupported)
         {
-            Span<short> output = scratch;
-            fixed (short* p = output)
+            fixed (short* ptr = &MemoryMarshal.GetReference(scratch))
             {
                 Vector128<byte> a0 = Sse2.ConvertScalarToVector128UInt32(a).AsByte();
                 Vector128<byte> b0 = Sse2.ConvertScalarToVector128UInt32(b).AsByte();
@@ -1489,8 +1488,8 @@ private static uint Select(uint a, uint b, uint c, Span<short> scratch)
                 Vector128<byte> pa = Sse2.UnpackLow(ac, Vector128<byte>.Zero); // |a - c|
                 Vector128<byte> pb = Sse2.UnpackLow(bc, Vector128<byte>.Zero); // |b - c|
                 Vector128<ushort> diff = Sse2.Subtract(pb.AsUInt16(), pa.AsUInt16());
-                Sse2.Store((ushort*)p, diff);
-                int paMinusPb = output[3] + output[2] + output[1] + output[0];
+                Sse2.Store((ushort*)ptr, diff);
+                int paMinusPb = ptr[3] + ptr[2] + ptr[1] + ptr[0];
                 return (paMinusPb <= 0) ? a : b;
             }
         }
diff --git a/src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs b/src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs
index 689c63f5b1..4113579637 100644
--- a/src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs
+++ b/src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs
@@ -353,8 +353,8 @@ private static void GetResidual(
         else
         {
 #pragma warning disable SA1503 // Braces should not be omitted
-            fixed (uint* currentRow = currentRowSpan)
-            fixed (uint* upperRow = upperRowSpan)
+            fixed (uint* currentRow = &MemoryMarshal.GetReference(currentRowSpan))
+            fixed (uint* upperRow = &MemoryMarshal.GetReference(upperRowSpan))
             {
                 for (int x = xStart; x < xEnd; x++)
                 {
@@ -664,9 +664,9 @@ private static void PredictBatch(
         Span<short> scratch)
     {
 #pragma warning disable SA1503 // Braces should not be omitted
-        fixed (uint* current = currentSpan)
-        fixed (uint* upper = upperSpan)
-        fixed (uint* outputFixed = outputSpan)
+        fixed (uint* current = &MemoryMarshal.GetReference(currentSpan))
+        fixed (uint* upper = &MemoryMarshal.GetReference(upperSpan))
+        fixed (uint* outputFixed = &MemoryMarshal.GetReference(outputSpan))
         {
             uint* output = outputFixed;
             if (xStart == 0)
diff --git a/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs b/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
index 1a8fcbafc9..735d0bf557 100644
--- a/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
+++ b/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
@@ -25,7 +25,7 @@ public static unsafe bool CheckNonOpaque(Span<Bgra32> row)
             ReadOnlySpan<byte> rowBytes = MemoryMarshal.AsBytes(row);
             int i = 0;
             int length = (row.Length * 4) - 3;
-            fixed (byte* src = rowBytes)
+            fixed (byte* src = &MemoryMarshal.GetReference(rowBytes))
             {
                 var alphaMaskVector256 = Vector256.Create(0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255);
                 Vector256<byte> all0x80Vector256 = Vector256.Create((byte)0x80).AsByte();
@@ -81,7 +81,7 @@ public static unsafe bool CheckNonOpaque(Span<Bgra32> row)
             ReadOnlySpan<byte> rowBytes = MemoryMarshal.AsBytes(row);
             int i = 0;
             int length = (row.Length * 4) - 3;
-            fixed (byte* src = rowBytes)
+            fixed (byte* src = &MemoryMarshal.GetReference(rowBytes))
             {
                 for (; i + 64 <= length; i += 64)
                 {
diff --git a/src/ImageSharp/Memory/Buffer2DExtensions.cs b/src/ImageSharp/Memory/Buffer2DExtensions.cs
index 2eb05ea935..31617c163b 100644
--- a/src/ImageSharp/Memory/Buffer2DExtensions.cs
+++ b/src/ImageSharp/Memory/Buffer2DExtensions.cs
@@ -50,7 +50,7 @@ internal static unsafe void DangerousCopyColumns<T>(
 
         Span<byte> span = MemoryMarshal.AsBytes(buffer.DangerousGetSingleMemory().Span);
 
-        fixed (byte* ptr = span)
+        fixed (byte* ptr = &MemoryMarshal.GetReference(span))
         {
             byte* basePtr = ptr;
             for (int y = 0; y < buffer.Height; y++)

From 45321948c837b675b6339dafaf7ac8d7fc5e9421 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?G=C3=BCnther=20Foidl?= <gue@korporal.at>
Date: Sun, 26 Mar 2023 14:19:26 +0200
Subject: [PATCH 2/6] Block8x8 set explicit size instead of (unused) fixed
 sized buffer field

---
 .../Formats/Jpeg/Components/Block8x8.cs           | 15 ++-------------
 1 file changed, 2 insertions(+), 13 deletions(-)

diff --git a/src/ImageSharp/Formats/Jpeg/Components/Block8x8.cs b/src/ImageSharp/Formats/Jpeg/Components/Block8x8.cs
index d119a18c8b..b417a8c814 100644
--- a/src/ImageSharp/Formats/Jpeg/Components/Block8x8.cs
+++ b/src/ImageSharp/Formats/Jpeg/Components/Block8x8.cs
@@ -15,25 +15,14 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
 /// 8x8 matrix of <see cref="short"/> coefficients.
 /// </summary>
 // ReSharper disable once InconsistentNaming
-[StructLayout(LayoutKind.Explicit)]
-internal unsafe partial struct Block8x8
+[StructLayout(LayoutKind.Explicit, Size = 2 * Size)]
+internal partial struct Block8x8
 {
     /// <summary>
     /// A number of scalar coefficients in a <see cref="Block8x8F"/>
     /// </summary>
     public const int Size = 64;
 
-#pragma warning disable IDE0051 // Remove unused private member
-    /// <summary>
-    /// A placeholder buffer so the actual struct occupies exactly 64 * 2 bytes.
-    /// </summary>
-    /// <remarks>
-    /// This is not used directly in the code.
-    /// </remarks>
-    [FieldOffset(0)]
-    private fixed short data[Size];
-#pragma warning restore IDE0051
-
     /// <summary>
     /// Gets or sets a <see cref="short"/> value at the given index
     /// </summary>

From a534328dc482683834c6901d93da61cf0d053724 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?G=C3=BCnther=20Foidl?= <gue@korporal.at>
Date: Sun, 26 Mar 2023 14:34:07 +0200
Subject: [PATCH 3/6] Optimized Block8x8F

---
 .../Formats/Jpeg/Components/Block8x8F.cs      | 25 +++++++------------
 .../Block8x8F_MultiplyInPlaceBlock.cs         |  4 +--
 .../Formats/Jpg/Block8x8FTests.cs             | 12 +++------
 .../ImageSharp.Tests/Formats/Jpg/DCTTests.cs  | 15 ++++-------
 ...ferenceImplementationsTests.AccurateDCT.cs |  3 +--
 ...plementationsTests.FastFloatingPointDCT.cs |  3 +--
 ...ceImplementations.LLM_FloatingPoint_DCT.cs |  6 ++---
 7 files changed, 23 insertions(+), 45 deletions(-)

diff --git a/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs b/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs
index 21971a8c7d..d432e82d24 100644
--- a/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs
+++ b/src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs
@@ -101,24 +101,17 @@ internal float this[nuint idx]
         set => this[((uint)y * 8) + (uint)x] = value;
     }
 
-    public static Block8x8F Load(Span<float> data)
-    {
-        Block8x8F result = default;
-        result.LoadFrom(data);
-        return result;
-    }
-
     /// <summary>
     /// Load raw 32bit floating point data from source.
     /// </summary>
-    /// <param name="source">Source</param>
+    /// <param name="data">Source</param>
     [MethodImpl(InliningOptions.ShortMethod)]
-    public void LoadFrom(Span<float> source)
+    public static Block8x8F Load(Span<float> data)
     {
-        ref byte s = ref Unsafe.As<float, byte>(ref MemoryMarshal.GetReference(source));
-        ref byte d = ref Unsafe.As<Block8x8F, byte>(ref this);
+        DebugGuard.MustBeGreaterThanOrEqualTo(data.Length, Size, "data is too small");
 
-        Unsafe.CopyBlock(ref d, ref s, Size * sizeof(float));
+        ref byte src = ref Unsafe.As<float, byte>(ref MemoryMarshal.GetReference(data));
+        return Unsafe.ReadUnaligned<Block8x8F>(ref src);
     }
 
     /// <summary>
@@ -144,10 +137,10 @@ public unsafe void LoadFrom(Span<int> source)
     [MethodImpl(InliningOptions.ShortMethod)]
     public unsafe void ScaledCopyTo(float[] dest)
     {
-        fixed (void* ptr = &this.V0L)
-        {
-            Marshal.Copy((IntPtr)ptr, dest, 0, Size);
-        }
+        DebugGuard.MustBeGreaterThanOrEqualTo(dest.Length, Size, "dest is too small");
+
+        ref byte destRef = ref Unsafe.As<float, byte>(ref MemoryMarshal.GetArrayDataReference(dest));
+        Unsafe.WriteUnaligned(ref destRef, this);
     }
 
     public float[] ToArray()
diff --git a/tests/ImageSharp.Benchmarks/Codecs/Jpeg/BlockOperations/Block8x8F_MultiplyInPlaceBlock.cs b/tests/ImageSharp.Benchmarks/Codecs/Jpeg/BlockOperations/Block8x8F_MultiplyInPlaceBlock.cs
index a5abeb3b66..722b095870 100644
--- a/tests/ImageSharp.Benchmarks/Codecs/Jpeg/BlockOperations/Block8x8F_MultiplyInPlaceBlock.cs
+++ b/tests/ImageSharp.Benchmarks/Codecs/Jpeg/BlockOperations/Block8x8F_MultiplyInPlaceBlock.cs
@@ -29,8 +29,6 @@ private static Block8x8F Create8x8FloatData()
             }
         }
 
-        var source = default(Block8x8F);
-        source.LoadFrom(result);
-        return source;
+        return Block8x8F.Load(result);
     }
 }
diff --git a/tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs b/tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs
index 3e4bcae6b7..cde9e776b2 100644
--- a/tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs
+++ b/tests/ImageSharp.Tests/Formats/Jpg/Block8x8FTests.cs
@@ -99,8 +99,7 @@ public void Load_Store_FloatArray()
             Times,
             () =>
             {
-                var b = default(Block8x8F);
-                b.LoadFrom(data);
+                Block8x8F b = Block8x8F.Load(data);
                 b.ScaledCopyTo(mirror);
             });
 
@@ -117,8 +116,7 @@ static void RunTest()
             float[] expected = Create8x8FloatData();
             ReferenceImplementations.Transpose8x8(expected);
 
-            var block8x8 = default(Block8x8F);
-            block8x8.LoadFrom(Create8x8FloatData());
+            Block8x8F block8x8 = Block8x8F.Load(Create8x8FloatData());
 
             block8x8.TransposeInplace();
 
@@ -153,9 +151,8 @@ private static float[] Create8x8ColorCropTestData()
     [Fact]
     public void NormalizeColors()
     {
-        var block = default(Block8x8F);
         float[] input = Create8x8ColorCropTestData();
-        block.LoadFrom(input);
+        Block8x8F block = Block8x8F.Load(input);
         this.Output.WriteLine("Input:");
         this.PrintLinearData(input);
 
@@ -242,8 +239,7 @@ public void RoundInto()
     {
         float[] data = Create8x8RandomFloatData(-1000, 1000);
 
-        var source = default(Block8x8F);
-        source.LoadFrom(data);
+        Block8x8F source = Block8x8F.Load(data);
         var dest = default(Block8x8);
 
         source.RoundInto(ref dest);
diff --git a/tests/ImageSharp.Tests/Formats/Jpg/DCTTests.cs b/tests/ImageSharp.Tests/Formats/Jpg/DCTTests.cs
index 5853ff37a5..5a1488c411 100644
--- a/tests/ImageSharp.Tests/Formats/Jpg/DCTTests.cs
+++ b/tests/ImageSharp.Tests/Formats/Jpg/DCTTests.cs
@@ -114,8 +114,7 @@ static void RunTest(string serialized)
                 int seed = FeatureTestRunner.Deserialize<int>(serialized);
 
                 Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed);
-                var srcBlock = default(Block8x8F);
-                srcBlock.LoadFrom(src);
+                Block8x8F srcBlock = Block8x8F.Load(src);
 
                 float[] expectedDest = new float[64];
                 float[] temp = new float[64];
@@ -162,8 +161,7 @@ static void RunTest(string serialized)
         public void TranformIDCT_4x4(int seed)
         {
             Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed, 4, 4);
-            var srcBlock = default(Block8x8F);
-            srcBlock.LoadFrom(src);
+            Block8x8F srcBlock = Block8x8F.Load(src);
 
             float[] expectedDest = new float[64];
             float[] temp = new float[64];
@@ -224,8 +222,7 @@ static void AssertScaledElementEquality(Span<float> expected, Span<float> actual
         public void TranformIDCT_2x2(int seed)
         {
             Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed, 2, 2);
-            var srcBlock = default(Block8x8F);
-            srcBlock.LoadFrom(src);
+            Block8x8F srcBlock = Block8x8F.Load(src);
 
             float[] expectedDest = new float[64];
             float[] temp = new float[64];
@@ -286,8 +283,7 @@ static void AssertScaledElementEquality(Span<float> expected, Span<float> actual
         public void TranformIDCT_1x1(int seed)
         {
             Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed, 1, 1);
-            var srcBlock = default(Block8x8F);
-            srcBlock.LoadFrom(src);
+            Block8x8F srcBlock = Block8x8F.Load(src);
 
             float[] expectedDest = new float[64];
             float[] temp = new float[64];
@@ -330,8 +326,7 @@ static void RunTest(string serialized)
                 int seed = FeatureTestRunner.Deserialize<int>(serialized);
 
                 Span<float> src = Create8x8RandomFloatData(MinInputValue, MaxInputValue, seed);
-                var block = default(Block8x8F);
-                block.LoadFrom(src);
+                Block8x8F block = Block8x8F.Load(src);
 
                 float[] expectedDest = new float[64];
                 float[] temp1 = new float[64];
diff --git a/tests/ImageSharp.Tests/Formats/Jpg/ReferenceImplementationsTests.AccurateDCT.cs b/tests/ImageSharp.Tests/Formats/Jpg/ReferenceImplementationsTests.AccurateDCT.cs
index cd93adefd8..c593a029ab 100644
--- a/tests/ImageSharp.Tests/Formats/Jpg/ReferenceImplementationsTests.AccurateDCT.cs
+++ b/tests/ImageSharp.Tests/Formats/Jpg/ReferenceImplementationsTests.AccurateDCT.cs
@@ -25,8 +25,7 @@ public void ForwardThenInverse(int seed)
         {
             float[] data = Create8x8RandomFloatData(-1000, 1000, seed);
 
-            var b0 = default(Block8x8F);
-            b0.LoadFrom(data);
+            Block8x8F b0 = Block8x8F.Load(data);
 
             Block8x8F b1 = ReferenceImplementations.AccurateDCT.TransformFDCT(ref b0);
             Block8x8F b2 = ReferenceImplementations.AccurateDCT.TransformIDCT(ref b1);
diff --git a/tests/ImageSharp.Tests/Formats/Jpg/ReferenceImplementationsTests.FastFloatingPointDCT.cs b/tests/ImageSharp.Tests/Formats/Jpg/ReferenceImplementationsTests.FastFloatingPointDCT.cs
index 46c6ee2da6..f5d7c159ba 100644
--- a/tests/ImageSharp.Tests/Formats/Jpg/ReferenceImplementationsTests.FastFloatingPointDCT.cs
+++ b/tests/ImageSharp.Tests/Formats/Jpg/ReferenceImplementationsTests.FastFloatingPointDCT.cs
@@ -70,8 +70,7 @@ public void LLM_FDCT_IsEquivalentTo_AccurateImplementation(int seed)
         {
             float[] floatData = Create8x8RandomFloatData(-1000, 1000);
 
-            Block8x8F source = default;
-            source.LoadFrom(floatData);
+            Block8x8F source = Block8x8F.Load(floatData);
 
             Block8x8F expected = ReferenceImplementations.AccurateDCT.TransformFDCT(ref source);
             Block8x8F actual = ReferenceImplementations.LLM_FloatingPoint_DCT.TransformFDCT_UpscaleBy8(ref source);
diff --git a/tests/ImageSharp.Tests/Formats/Jpg/Utils/ReferenceImplementations.LLM_FloatingPoint_DCT.cs b/tests/ImageSharp.Tests/Formats/Jpg/Utils/ReferenceImplementations.LLM_FloatingPoint_DCT.cs
index 9e0c62d139..0d5f3114d1 100644
--- a/tests/ImageSharp.Tests/Formats/Jpg/Utils/ReferenceImplementations.LLM_FloatingPoint_DCT.cs
+++ b/tests/ImageSharp.Tests/Formats/Jpg/Utils/ReferenceImplementations.LLM_FloatingPoint_DCT.cs
@@ -36,8 +36,7 @@ public static Block8x8F TransformIDCT(ref Block8x8F source)
             float[] temp = new float[64];
 
             IDCT2D_llm(s, d, temp);
-            Block8x8F result = default;
-            result.LoadFrom(d);
+            Block8x8F result = Block8x8F.Load(d);
             return result;
         }
 
@@ -49,8 +48,7 @@ public static Block8x8F TransformFDCT_UpscaleBy8(ref Block8x8F source)
             float[] temp = new float[64];
 
             FDCT2D_llm(s, d, temp);
-            Block8x8F result = default;
-            result.LoadFrom(d);
+            Block8x8F result = Block8x8F.Load(d);
             return result;
         }
 

From f348d704efac5bcb001cd8ba40b12d1cc7329d8c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?G=C3=BCnther=20Foidl?= <gue@korporal.at>
Date: Mon, 27 Mar 2023 18:34:02 +0200
Subject: [PATCH 4/6] Block8x8 Load and CopyTo simplified

Similar to a534328dc482683834c6901d93da61cf0d053724
---
 .../Formats/Jpeg/Components/Block8x8.cs       | 27 ++++++-------------
 .../Formats/Jpg/Block8x8Tests.cs              |  3 +--
 2 files changed, 9 insertions(+), 21 deletions(-)

diff --git a/src/ImageSharp/Formats/Jpeg/Components/Block8x8.cs b/src/ImageSharp/Formats/Jpeg/Components/Block8x8.cs
index b417a8c814..01d112bd6f 100644
--- a/src/ImageSharp/Formats/Jpeg/Components/Block8x8.cs
+++ b/src/ImageSharp/Formats/Jpeg/Components/Block8x8.cs
@@ -63,9 +63,10 @@ public short this[int idx]
 
     public static Block8x8 Load(Span<short> data)
     {
-        Unsafe.SkipInit(out Block8x8 result);
-        result.LoadFrom(data);
-        return result;
+        DebugGuard.MustBeGreaterThanOrEqualTo(data.Length, Size, "data is too small");
+
+        ref byte src = ref Unsafe.As<short, byte>(ref MemoryMarshal.GetReference(data));
+        return Unsafe.ReadUnaligned<Block8x8>(ref src);
     }
 
     /// <summary>
@@ -93,9 +94,10 @@ public short[] ToArray()
     /// </summary>
     public void CopyTo(Span<short> destination)
     {
-        ref byte selfRef = ref Unsafe.As<Block8x8, byte>(ref this);
-        ref byte destRef = ref MemoryMarshal.GetReference(MemoryMarshal.Cast<short, byte>(destination));
-        Unsafe.CopyBlockUnaligned(ref destRef, ref selfRef, Size * sizeof(short));
+        DebugGuard.MustBeGreaterThanOrEqualTo(destination.Length, Size, "destination is too small");
+
+        ref byte destRef = ref Unsafe.As<short, byte>(ref MemoryMarshal.GetReference(destination));
+        Unsafe.WriteUnaligned(ref destRef, this);
     }
 
     /// <summary>
@@ -124,19 +126,6 @@ public void LoadFrom(ReadOnlySpan<byte> source)
         }
     }
 
-    /// <summary>
-    /// Load raw 16bit integers from source.
-    /// </summary>
-    /// <param name="source">Source</param>
-    [MethodImpl(InliningOptions.ShortMethod)]
-    public void LoadFrom(Span<short> source)
-    {
-        ref byte sourceRef = ref Unsafe.As<short, byte>(ref MemoryMarshal.GetReference(source));
-        ref byte destRef = ref Unsafe.As<Block8x8, byte>(ref this);
-
-        Unsafe.CopyBlockUnaligned(ref destRef, ref sourceRef, Size * sizeof(short));
-    }
-
     /// <summary>
     /// Cast and copy <see cref="Size"/> <see cref="int"/>-s from the beginning of 'source' span.
     /// </summary>
diff --git a/tests/ImageSharp.Tests/Formats/Jpg/Block8x8Tests.cs b/tests/ImageSharp.Tests/Formats/Jpg/Block8x8Tests.cs
index 798ea30407..b5d364dd38 100644
--- a/tests/ImageSharp.Tests/Formats/Jpg/Block8x8Tests.cs
+++ b/tests/ImageSharp.Tests/Formats/Jpg/Block8x8Tests.cs
@@ -269,8 +269,7 @@ static void RunTest()
             short[] expected = Create8x8ShortData();
             ReferenceImplementations.Transpose8x8(expected);
 
-            var block8x8 = default(Block8x8);
-            block8x8.LoadFrom(Create8x8ShortData());
+            Block8x8 block8x8 = Block8x8.Load(Create8x8ShortData());
 
             block8x8.TransposeInplace();
 

From 4b84aecb94d37e934051273e2487ebe798424c0b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?G=C3=BCnther=20Foidl?= <gue@korporal.at>
Date: Mon, 27 Mar 2023 19:09:25 +0200
Subject: [PATCH 5/6] Reverted some pinning changes to have less verbose C#
 code

---
 .../Common/Helpers/SimdUtils.HwIntrinsics.cs           |  2 +-
 src/ImageSharp/Compression/Zlib/Adler32.cs             |  2 +-
 src/ImageSharp/Compression/Zlib/Crc32.cs               |  6 +++---
 src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs  |  2 +-
 .../Formats/Webp/Lossless/PredictorEncoder.cs          | 10 +++++-----
 src/ImageSharp/Formats/Webp/WebpCommonUtils.cs         |  2 +-
 src/ImageSharp/Memory/Buffer2DExtensions.cs            |  2 +-
 7 files changed, 13 insertions(+), 13 deletions(-)

diff --git a/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs b/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs
index fd8a61350f..8ba1b09125 100644
--- a/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs
+++ b/src/ImageSharp/Common/Helpers/SimdUtils.HwIntrinsics.cs
@@ -673,7 +673,7 @@ internal static unsafe void ByteToNormalizedFloat(
             ReadOnlySpan<byte> source,
             Span<float> dest)
         {
-            fixed (byte* sourceBase = &MemoryMarshal.GetReference(source))
+            fixed (byte* sourceBase = source)
             {
                 if (Avx2.IsSupported)
                 {
diff --git a/src/ImageSharp/Compression/Zlib/Adler32.cs b/src/ImageSharp/Compression/Zlib/Adler32.cs
index 3ecdf81539..dd8217541a 100644
--- a/src/ImageSharp/Compression/Zlib/Adler32.cs
+++ b/src/ImageSharp/Compression/Zlib/Adler32.cs
@@ -387,7 +387,7 @@ private static unsafe uint CalculateScalar(uint adler, ReadOnlySpan<byte> buffer
         uint s1 = adler & 0xFFFF;
         uint s2 = (adler >> 16) & 0xFFFF;
 
-        fixed (byte* bufferPtr = &MemoryMarshal.GetReference(buffer))
+        fixed (byte* bufferPtr = buffer)
         {
             byte* localBufferPtr = bufferPtr;
             uint length = (uint)buffer.Length;
diff --git a/src/ImageSharp/Compression/Zlib/Crc32.cs b/src/ImageSharp/Compression/Zlib/Crc32.cs
index e21621ab74..2d0a09bd4c 100644
--- a/src/ImageSharp/Compression/Zlib/Crc32.cs
+++ b/src/ImageSharp/Compression/Zlib/Crc32.cs
@@ -81,7 +81,7 @@ private static unsafe uint CalculateSse(uint crc, ReadOnlySpan<byte> buffer)
         int chunksize = buffer.Length & ~ChunksizeMask;
         int length = chunksize;
 
-        fixed (byte* bufferPtr = &MemoryMarshal.GetReference(buffer))
+        fixed (byte* bufferPtr = buffer)
         {
             fixed (ulong* k05PolyPtr = K05Poly)
             {
@@ -201,7 +201,7 @@ private static unsafe uint CalculateSse(uint crc, ReadOnlySpan<byte> buffer)
     [MethodImpl(InliningOptions.HotPath | InliningOptions.ShortMethod)]
     private static unsafe uint CalculateArm(uint crc, ReadOnlySpan<byte> buffer)
     {
-        fixed (byte* bufferPtr = &MemoryMarshal.GetReference(buffer))
+        fixed (byte* bufferPtr = buffer)
         {
             byte* localBufferPtr = bufferPtr;
             int len = buffer.Length;
@@ -248,7 +248,7 @@ private static unsafe uint CalculateArm(uint crc, ReadOnlySpan<byte> buffer)
     [MethodImpl(InliningOptions.HotPath | InliningOptions.ShortMethod)]
     private static unsafe uint CalculateArm64(uint crc, ReadOnlySpan<byte> buffer)
     {
-        fixed (byte* bufferPtr = &MemoryMarshal.GetReference(buffer))
+        fixed (byte* bufferPtr = buffer)
         {
             byte* localBufferPtr = bufferPtr;
             int len = buffer.Length;
diff --git a/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs b/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
index ab6e4cfccf..024adb7c23 100644
--- a/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
+++ b/src/ImageSharp/Formats/Webp/Lossless/LosslessUtils.cs
@@ -569,7 +569,7 @@ public static void PredictorInverseTransform(
         Span<uint> pixelData,
         Span<uint> outputSpan)
     {
-        fixed (uint* inputFixed = &MemoryMarshal.GetReference(pixelData))
+        fixed (uint* inputFixed = pixelData)
         {
             fixed (uint* outputFixed = outputSpan)
             {
diff --git a/src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs b/src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs
index 4113579637..689c63f5b1 100644
--- a/src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs
+++ b/src/ImageSharp/Formats/Webp/Lossless/PredictorEncoder.cs
@@ -353,8 +353,8 @@ private static void GetResidual(
         else
         {
 #pragma warning disable SA1503 // Braces should not be omitted
-            fixed (uint* currentRow = &MemoryMarshal.GetReference(currentRowSpan))
-            fixed (uint* upperRow = &MemoryMarshal.GetReference(upperRowSpan))
+            fixed (uint* currentRow = currentRowSpan)
+            fixed (uint* upperRow = upperRowSpan)
             {
                 for (int x = xStart; x < xEnd; x++)
                 {
@@ -664,9 +664,9 @@ private static void PredictBatch(
         Span<short> scratch)
     {
 #pragma warning disable SA1503 // Braces should not be omitted
-        fixed (uint* current = &MemoryMarshal.GetReference(currentSpan))
-        fixed (uint* upper = &MemoryMarshal.GetReference(upperSpan))
-        fixed (uint* outputFixed = &MemoryMarshal.GetReference(outputSpan))
+        fixed (uint* current = currentSpan)
+        fixed (uint* upper = upperSpan)
+        fixed (uint* outputFixed = outputSpan)
         {
             uint* output = outputFixed;
             if (xStart == 0)
diff --git a/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs b/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
index 735d0bf557..4a7dc74ca7 100644
--- a/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
+++ b/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
@@ -81,7 +81,7 @@ public static unsafe bool CheckNonOpaque(Span<Bgra32> row)
             ReadOnlySpan<byte> rowBytes = MemoryMarshal.AsBytes(row);
             int i = 0;
             int length = (row.Length * 4) - 3;
-            fixed (byte* src = &MemoryMarshal.GetReference(rowBytes))
+            fixed (byte* src = rowBytes)
             {
                 for (; i + 64 <= length; i += 64)
                 {
diff --git a/src/ImageSharp/Memory/Buffer2DExtensions.cs b/src/ImageSharp/Memory/Buffer2DExtensions.cs
index 31617c163b..2eb05ea935 100644
--- a/src/ImageSharp/Memory/Buffer2DExtensions.cs
+++ b/src/ImageSharp/Memory/Buffer2DExtensions.cs
@@ -50,7 +50,7 @@ internal static unsafe void DangerousCopyColumns<T>(
 
         Span<byte> span = MemoryMarshal.AsBytes(buffer.DangerousGetSingleMemory().Span);
 
-        fixed (byte* ptr = &MemoryMarshal.GetReference(span))
+        fixed (byte* ptr = span)
         {
             byte* basePtr = ptr;
             for (int y = 0; y < buffer.Height; y++)

From 66b2f327bed811d7abce65921148d797ee4fdbd3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?G=C3=BCnther=20Foidl?= <gue@korporal.at>
Date: Mon, 27 Mar 2023 19:19:02 +0200
Subject: [PATCH 6/6] Reverted Webp/WebpCommonUtils fixed-verbosity

---
 src/ImageSharp/Formats/Webp/WebpCommonUtils.cs | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs b/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
index 4a7dc74ca7..1a8fcbafc9 100644
--- a/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
+++ b/src/ImageSharp/Formats/Webp/WebpCommonUtils.cs
@@ -25,7 +25,7 @@ public static unsafe bool CheckNonOpaque(Span<Bgra32> row)
             ReadOnlySpan<byte> rowBytes = MemoryMarshal.AsBytes(row);
             int i = 0;
             int length = (row.Length * 4) - 3;
-            fixed (byte* src = &MemoryMarshal.GetReference(rowBytes))
+            fixed (byte* src = rowBytes)
             {
                 var alphaMaskVector256 = Vector256.Create(0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255);
                 Vector256<byte> all0x80Vector256 = Vector256.Create((byte)0x80).AsByte();