Used unsigned division for vector sizes to get better codegen

Cf. https://sharplab.io/#v2:EYLgxg9gTgpgtADwGwBYA0AXEBDAzgWwB8ABAJgEYBYAKGIAYACY8gOgCUBXAOwwEt8YLAJI8ovLrl5hcAbho1iAZiakGAYQYBvGg11NlXcRgYBZcgAojDADYwuAcwwALAJQMAvAD4bdx04YA9AwAajBgGNDkpAAcADwAZtYQ2BieLGoQ3Bhy1Hr6DIY8pqSWRbYOzm5eDOaFGC7mHEYu5X6BIWERUFFxicmp6Zk8OQC+QA=

src/ImageSharp/Common/Helpers/Numerics.cs

@@ -55,6 +55,12 @@ internal static class Numerics
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static int Modulo4(int x) => x & 3;
 
+    /// <summary>
+    /// Calculates <paramref name="x"/> % 4
+    /// </summary>
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static nint Modulo4(nint x) => x & 3;
+
     /// <summary>
     /// Calculates <paramref name="x"/> % 8
     /// </summary>
@@ -430,9 +436,9 @@ internal static class Numerics
         var vmin = new Vector<T>(min);
         var vmax = new Vector<T>(max);
 
-        int n = span.Length / Vector<T>.Count;
-        int m = Modulo4(n);
-        int u = n - m;
+        nint n = (nint)(uint)span.Length / Vector<T>.Count;
+        nint m = Modulo4(n);
+        nint u = n - m;
 
         ref Vector<T> vs0 = ref Unsafe.As<T, Vector<T>>(ref MemoryMarshal.GetReference(span));
         ref Vector<T> vs1 = ref Unsafe.Add(ref vs0, 1);

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

@@ -97,12 +97,12 @@ internal static partial class SimdUtils
         {
             VerifySpanInput(source, dest, Vector<byte>.Count);
 
-            int n = dest.Length / Vector<byte>.Count;
+            nint n = (nint)(uint)dest.Length / Vector<byte>.Count;
 
             ref Vector<byte> sourceBase = ref Unsafe.As<byte, Vector<byte>>(ref MemoryMarshal.GetReference(source));
             ref Vector<float> destBase = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(dest));
 
-            for (int i = 0; i < n; i++)
+            for (nint i = 0; i < n; i++)
             {
                 Vector<byte> b = Unsafe.Add(ref sourceBase, i);
 
@@ -132,13 +132,13 @@ internal static partial class SimdUtils
         {
             VerifySpanInput(source, dest, Vector<byte>.Count);
 
-            int n = dest.Length / Vector<byte>.Count;
+            nint n = (nint)(uint)dest.Length / Vector<byte>.Count;
 
             ref Vector<float> sourceBase =
                 ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(source));
             ref Vector<byte> destBase = ref Unsafe.As<byte, Vector<byte>>(ref MemoryMarshal.GetReference(dest));
 
-            for (int i = 0; i < n; i++)
+            for (nint i = 0; i < n; i++)
             {
                 ref Vector<float> s = ref Unsafe.Add(ref sourceBase, i * 4);
 

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

@@ -221,11 +221,11 @@ internal static partial class SimdUtils
                 ref Vector256<float> destBase =
                     ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(dest));
 
-                int n = dest.Length / Vector256<float>.Count;
-                int m = Numerics.Modulo4(n);
-                int u = n - m;
+                nint n = (nint)(uint)dest.Length / Vector256<float>.Count;
+                nint m = Numerics.Modulo4(n);
+                nint u = n - m;
 
-                for (int i = 0; i < u; i += 4)
+                for (nint i = 0; i < u; i += 4)
                 {
                     ref Vector256<float> vd0 = ref Unsafe.Add(ref destBase, i);
                     ref Vector256<float> vs0 = ref Unsafe.Add(ref sourceBase, i);
@@ -238,7 +238,7 @@ internal static partial class SimdUtils
 
                 if (m > 0)
                 {
-                    for (int i = u; i < n; i++)
+                    for (nint i = u; i < n; i++)
                     {
                         Unsafe.Add(ref destBase, i) = Avx.Permute(Unsafe.Add(ref sourceBase, i), control);
                     }
@@ -253,11 +253,11 @@ internal static partial class SimdUtils
                 ref Vector128<float> destBase =
                     ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(dest));
 
-                int n = dest.Length / Vector128<float>.Count;
-                int m = Numerics.Modulo4(n);
-                int u = n - m;
+                nint n = (nint)(uint)dest.Length / Vector128<float>.Count;
+                nint m = Numerics.Modulo4(n);
+                nint u = n - m;
 
-                for (int i = 0; i < u; i += 4)
+                for (nint i = 0; i < u; i += 4)
                 {
                     ref Vector128<float> vd0 = ref Unsafe.Add(ref destBase, i);
                     ref Vector128<float> vs0 = ref Unsafe.Add(ref sourceBase, i);
@@ -276,7 +276,7 @@ internal static partial class SimdUtils
 
                 if (m > 0)
                 {
-                    for (int i = u; i < n; i++)
+                    for (nint i = u; i < n; i++)
                     {
                         Vector128<float> vs = Unsafe.Add(ref sourceBase, i);
                         Unsafe.Add(ref destBase, i) = Sse.Shuffle(vs, vs, control);
@@ -306,11 +306,11 @@ internal static partial class SimdUtils
                 ref Vector256<byte> destBase =
                     ref Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(dest));
 
-                int n = dest.Length / Vector256<byte>.Count;
-                int m = Numerics.Modulo4(n);
-                int u = n - m;
+                nint n = (nint)(uint)dest.Length / Vector256<byte>.Count;
+                nint m = Numerics.Modulo4(n);
+                nint u = n - m;
 
-                for (int i = 0; i < u; i += 4)
+                for (nint i = 0; i < u; i += 4)
                 {
                     ref Vector256<byte> vs0 = ref Unsafe.Add(ref sourceBase, i);
                     ref Vector256<byte> vd0 = ref Unsafe.Add(ref destBase, i);
@@ -323,7 +323,7 @@ internal static partial class SimdUtils
 
                 if (m > 0)
                 {
-                    for (int i = u; i < n; i++)
+                    for (nint i = u; i < n; i++)
                     {
                         Unsafe.Add(ref destBase, i) = Avx2.Shuffle(Unsafe.Add(ref sourceBase, i), vshuffle);
                     }
@@ -342,11 +342,11 @@ internal static partial class SimdUtils
                 ref Vector128<byte> destBase =
                     ref Unsafe.As<byte, Vector128<byte>>(ref MemoryMarshal.GetReference(dest));
 
-                int n = dest.Length / Vector128<byte>.Count;
-                int m = Numerics.Modulo4(n);
-                int u = n - m;
+                nint n = (nint)(uint)dest.Length / Vector128<byte>.Count;
+                nint m = Numerics.Modulo4(n);
+                nint u = n - m;
 
-                for (int i = 0; i < u; i += 4)
+                for (nint i = 0; i < u; i += 4)
                 {
                     ref Vector128<byte> vs0 = ref Unsafe.Add(ref sourceBase, i);
                     ref Vector128<byte> vd0 = ref Unsafe.Add(ref destBase, i);
@@ -359,7 +359,7 @@ internal static partial class SimdUtils
 
                 if (m > 0)
                 {
-                    for (int i = u; i < n; i++)
+                    for (nint i = u; i < n; i++)
                     {
                         Unsafe.Add(ref destBase, i) = Ssse3.Shuffle(Unsafe.Add(ref sourceBase, i), vshuffle);
                     }
@@ -391,9 +391,9 @@ internal static partial class SimdUtils
                 ref Vector128<byte> destBase =
                     ref Unsafe.As<byte, Vector128<byte>>(ref MemoryMarshal.GetReference(dest));
 
-                int n = source.Length / Vector128<byte>.Count;
+                nint n = (nint)(uint)source.Length / Vector128<byte>.Count;
 
-                for (int i = 0; i < n; i += 3)
+                for (nint i = 0; i < n; i += 3)
                 {
                     ref Vector128<byte> vs = ref Unsafe.Add(ref sourceBase, i);
 
@@ -454,9 +454,9 @@ internal static partial class SimdUtils
                 ref Vector128<byte> destBase =
                     ref Unsafe.As<byte, Vector128<byte>>(ref MemoryMarshal.GetReference(dest));
 
-                int n = source.Length / Vector128<byte>.Count;
+                nint n = (nint)(uint)source.Length / Vector128<byte>.Count;
 
-                for (int i = 0, j = 0; i < n; i += 3, j += 4)
+                for (nint i = 0, j = 0; i < n; i += 3, j += 4)
                 {
                     ref Vector128<byte> v0 = ref Unsafe.Add(ref sourceBase, i);
                     Vector128<byte> v1 = Unsafe.Add(ref v0, 1);
@@ -498,9 +498,9 @@ internal static partial class SimdUtils
                 ref Vector128<byte> destBase =
                     ref Unsafe.As<byte, Vector128<byte>>(ref MemoryMarshal.GetReference(dest));
 
-                int n = source.Length / Vector128<byte>.Count;
+                nint n = (nint)(uint)source.Length / Vector128<byte>.Count;
 
-                for (int i = 0, j = 0; i < n; i += 4, j += 3)
+                for (nint i = 0, j = 0; i < n; i += 4, j += 3)
                 {
                     ref Vector128<byte> vs = ref Unsafe.Add(ref sourceBase, i);
 
@@ -650,16 +650,16 @@ internal static partial class SimdUtils
                 {
                     VerifySpanInput(source, dest, Vector256<byte>.Count);
 
-                    int n = dest.Length / Vector256<byte>.Count;
+                    nint n = (nint)(uint)dest.Length / Vector256<byte>.Count;
 
                     ref Vector256<float> destBase =
                         ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(dest));
 
                     var scale = Vector256.Create(1 / (float)byte.MaxValue);
 
-                    for (int i = 0; i < n; i++)
+                    for (nint i = 0; i < n; i++)
                     {
-                        int si = Vector256<byte>.Count * i;
+                        nint si = Vector256<byte>.Count * i;
                         Vector256<int> i0 = Avx2.ConvertToVector256Int32(sourceBase + si);
                         Vector256<int> i1 = Avx2.ConvertToVector256Int32(sourceBase + si + Vector256<int>.Count);
                         Vector256<int> i2 = Avx2.ConvertToVector256Int32(sourceBase + si + (Vector256<int>.Count * 2));
@@ -683,7 +683,7 @@ internal static partial class SimdUtils
                     // Sse
                     VerifySpanInput(source, dest, Vector128<byte>.Count);
 
-                    int n = dest.Length / Vector128<byte>.Count;
+                    nint n = (nint)(uint)dest.Length / Vector128<byte>.Count;
 
                     ref Vector128<float> destBase =
                         ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(dest));
@@ -691,9 +691,9 @@ internal static partial class SimdUtils
                     var scale = Vector128.Create(1 / (float)byte.MaxValue);
                     Vector128<byte> zero = Vector128<byte>.Zero;
 
-                    for (int i = 0; i < n; i++)
+                    for (nint i = 0; i < n; i++)
                     {
-                        int si = Vector128<byte>.Count * i;
+                        nint si = Vector128<byte>.Count * i;
 
                         Vector128<int> i0, i1, i2, i3;
                         if (Sse41.IsSupported)
@@ -782,7 +782,7 @@ internal static partial class SimdUtils
             {
                 VerifySpanInput(source, dest, Vector256<byte>.Count);
 
-                int n = dest.Length / Vector256<byte>.Count;
+                nint n = (nint)(uint)dest.Length / Vector256<byte>.Count;
 
                 ref Vector256<float> sourceBase =
                     ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(source));
@@ -794,7 +794,7 @@ internal static partial class SimdUtils
                 ref byte maskBase = ref MemoryMarshal.GetReference(PermuteMaskDeinterleave8x32);
                 Vector256<int> mask = Unsafe.As<byte, Vector256<int>>(ref maskBase);
 
-                for (int i = 0; i < n; i++)
+                for (nint i = 0; i < n; i++)
                 {
                     ref Vector256<float> s = ref Unsafe.Add(ref sourceBase, i * 4);
 
@@ -821,7 +821,7 @@ internal static partial class SimdUtils
                 // Sse
                 VerifySpanInput(source, dest, Vector128<byte>.Count);
 
-                int n = dest.Length / Vector128<byte>.Count;
+                nint n = (nint)(uint)dest.Length / Vector128<byte>.Count;
 
                 ref Vector128<float> sourceBase =
                     ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(source));
@@ -831,7 +831,7 @@ internal static partial class SimdUtils
 
                 var scale = Vector128.Create((float)byte.MaxValue);
 
-                for (int i = 0; i < n; i++)
+                for (nint i = 0; i < n; i++)
                 {
                     ref Vector128<float> s = ref Unsafe.Add(ref sourceBase, i * 4);
 
@@ -864,7 +864,7 @@ internal static partial class SimdUtils
             ref Vector256<byte> bBase = ref Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(blueChannel));
             ref byte dBase = ref Unsafe.As<Rgb24, byte>(ref MemoryMarshal.GetReference(destination));
 
-            int count = redChannel.Length / Vector256<byte>.Count;
+            nint count = (nint)(uint)redChannel.Length / Vector256<byte>.Count;
 
             ref byte control1Bytes = ref MemoryMarshal.GetReference(PermuteMaskEvenOdd8x32);
             Vector256<uint> control1 = Unsafe.As<byte, Vector256<uint>>(ref control1Bytes);
@@ -875,7 +875,7 @@ internal static partial class SimdUtils
 
             Vector256<byte> shuffleAlpha = Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(ShuffleMaskShiftAlpha));
 
-            for (int i = 0; i < count; i++)
+            for (nint i = 0; i < count; i++)
             {
                 Vector256<byte> r0 = Unsafe.Add(ref rBase, i);
                 Vector256<byte> g0 = Unsafe.Add(ref gBase, i);
@@ -918,7 +918,7 @@ internal static partial class SimdUtils
                 Unsafe.As<byte, Vector256<byte>>(ref d4) = rgb4;
             }
 
-            int slice = count * Vector256<byte>.Count;
+            int slice = (int)count * Vector256<byte>.Count;
             redChannel = redChannel[slice..];
             greenChannel = greenChannel[slice..];
             blueChannel = blueChannel[slice..];
@@ -936,12 +936,12 @@ internal static partial class SimdUtils
             ref Vector256<byte> bBase = ref Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(blueChannel));
             ref Vector256<byte> dBase = ref Unsafe.As<Rgba32, Vector256<byte>>(ref MemoryMarshal.GetReference(destination));
 
-            int count = redChannel.Length / Vector256<byte>.Count;
+            nint count = (nint)(uint)redChannel.Length / Vector256<byte>.Count;
             ref byte control1Bytes = ref MemoryMarshal.GetReference(PermuteMaskEvenOdd8x32);
             Vector256<uint> control1 = Unsafe.As<byte, Vector256<uint>>(ref control1Bytes);
             var a = Vector256.Create((byte)255);
 
-            for (int i = 0; i < count; i++)
+            for (nint i = 0; i < count; i++)
             {
                 Vector256<byte> r0 = Unsafe.Add(ref rBase, i);
                 Vector256<byte> g0 = Unsafe.Add(ref gBase, i);
@@ -970,7 +970,7 @@ internal static partial class SimdUtils
                 Unsafe.Add(ref d0, 3) = rgb4;
             }
 
-            int slice = count * Vector256<byte>.Count;
+            int slice = (int)count * Vector256<byte>.Count;
             redChannel = redChannel[slice..];
             greenChannel = greenChannel[slice..];
             blueChannel = blueChannel[slice..];

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

@@ -425,7 +425,7 @@ internal partial struct Block8x8F : IEquatable<Block8x8F>
             Vector256<int> targetVector = Vector256.Create(value);
             ref Vector256<float> blockStride = ref this.V0;
 
-            for (int i = 0; i < RowCount; i++)
+            for (nint i = 0; i < RowCount; i++)
             {
                 Vector256<int> areEqual = Avx2.CompareEqual(Avx.ConvertToVector256Int32WithTruncation(Unsafe.Add(ref this.V0, i)), targetVector);
                 if (Avx2.MoveMask(areEqual.AsByte()) != equalityMask)
@@ -439,7 +439,7 @@ internal partial struct Block8x8F : IEquatable<Block8x8F>
 
         ref float scalars = ref Unsafe.As<Block8x8F, float>(ref this);
 
-        for (int i = 0; i < Size; i++)
+        for (nint i = 0; i < Size; i++)
         {
             if ((int)Unsafe.Add(ref scalars, i) != value)
             {

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

@@ -32,7 +32,7 @@ internal abstract partial class JpegColorConverterBase
             // Used for the color conversion
             var scale = Vector256.Create(1 / (this.MaximumValue * this.MaximumValue));
 
-            nint n = values.Component0.Length / Vector256<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector256<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 ref Vector256<float> c = ref Unsafe.Add(ref c0Base, i);
@@ -71,7 +71,7 @@ internal abstract partial class JpegColorConverterBase
 
             var scale = Vector256.Create(maxValue);
 
-            nint n = values.Component0.Length / Vector256<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector256<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 Vector256<float> ctmp = Avx.Subtract(scale, Unsafe.Add(ref srcR, i));

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

@@ -30,7 +30,7 @@ internal abstract partial class JpegColorConverterBase
 
             var scale = new Vector<float>(1 / (this.MaximumValue * this.MaximumValue));
 
-            nint n = values.Component0.Length / Vector<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 ref Vector<float> c = ref Unsafe.Add(ref cBase, i);
@@ -78,7 +78,7 @@ internal abstract partial class JpegColorConverterBase
             // Used for the color conversion
             var scale = new Vector<float>(maxValue);
 
-            nint n = values.Component0.Length / Vector<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 Vector<float> ctmp = scale - Unsafe.Add(ref srcR, i);

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

@@ -27,7 +27,7 @@ internal abstract partial class JpegColorConverterBase
             // Used for the color conversion
             var scale = Vector256.Create(1 / this.MaximumValue);
 
-            nint n = values.Component0.Length / Vector256<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector256<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 ref Vector256<float> c0 = ref Unsafe.Add(ref c0Base, i);
@@ -53,7 +53,7 @@ internal abstract partial class JpegColorConverterBase
             var f0587 = Vector256.Create(0.587f);
             var f0114 = Vector256.Create(0.114f);
 
-            nint n = values.Component0.Length / Vector256<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector256<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 ref Vector256<float> r = ref Unsafe.Add(ref srcRed, i);

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

@@ -24,7 +24,7 @@ internal abstract partial class JpegColorConverterBase
 
             var scale = new Vector<float>(1 / this.MaximumValue);
 
-            nint n = values.Component0.Length / Vector<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 ref Vector<float> c0 = ref Unsafe.Add(ref cBase, i);
@@ -53,7 +53,7 @@ internal abstract partial class JpegColorConverterBase
             var gMult = new Vector<float>(0.587f);
             var bMult = new Vector<float>(0.114f);
 
-            nint n = values.Component0.Length / Vector<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 Vector<float> r = Unsafe.Add(ref srcR, i);

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

@@ -29,7 +29,7 @@ internal abstract partial class JpegColorConverterBase
 
             // Used for the color conversion
             var scale = Vector256.Create(1 / this.MaximumValue);
-            nint n = values.Component0.Length / Vector256<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector256<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 ref Vector256<float> r = ref Unsafe.Add(ref rBase, i);

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

@@ -28,7 +28,7 @@ internal abstract partial class JpegColorConverterBase
 
             var scale = new Vector<float>(1 / this.MaximumValue);
 
-            nint n = values.Component0.Length / Vector<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 ref Vector<float> r = ref Unsafe.Add(ref rBase, i);

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

@@ -38,7 +38,7 @@ internal abstract partial class JpegColorConverterBase
             var bCbMult = Vector256.Create(YCbCrScalar.BCbMult);
 
             // Walking 8 elements at one step:
-            nint n = values.Component0.Length / Vector256<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector256<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 // y = yVals[i];
@@ -98,7 +98,7 @@ internal abstract partial class JpegColorConverterBase
             var fn0081312F = Vector256.Create(-0.081312F);
             var f05 = Vector256.Create(0.5f);
 
-            nint n = values.Component0.Length / Vector256<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector256<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 Vector256<float> r = Unsafe.Add(ref srcR, i);

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

@@ -35,7 +35,7 @@ internal abstract partial class JpegColorConverterBase
             var gCrMult = new Vector<float>(-YCbCrScalar.GCrMult);
             var bCbMult = new Vector<float>(YCbCrScalar.BCbMult);
 
-            nint n = values.Component0.Length / Vector<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 // y = yVals[i];
@@ -103,7 +103,7 @@ internal abstract partial class JpegColorConverterBase
             var gCrMult = new Vector<float>(0.418688f);
             var bCrMult = new Vector<float>(0.081312f);
 
-            nint n = values.Component0.Length / Vector<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 Vector<float> r = Unsafe.Add(ref srcR, i);

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

@@ -40,7 +40,7 @@ internal abstract partial class JpegColorConverterBase
             var bCbMult = Vector256.Create(YCbCrScalar.BCbMult);
 
             // Walking 8 elements at one step:
-            nint n = values.Component0.Length / Vector256<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector256<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 // y = yVals[i];
@@ -109,7 +109,7 @@ internal abstract partial class JpegColorConverterBase
             var fn0081312F = Vector256.Create(-0.081312F);
             var f05 = Vector256.Create(0.5f);
 
-            nint n = values.Component0.Length / Vector256<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector256<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 Vector256<float> r = Avx.Subtract(maxSampleValue, Unsafe.Add(ref srcR, i));

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

@@ -36,7 +36,7 @@ internal abstract partial class JpegColorConverterBase
             var gCrMult = new Vector<float>(-YCbCrScalar.GCrMult);
             var bCbMult = new Vector<float>(YCbCrScalar.BCbMult);
 
-            nint n = values.Component0.Length / Vector<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 // y = yVals[i];
@@ -107,7 +107,7 @@ internal abstract partial class JpegColorConverterBase
             var gCrMult = new Vector<float>(0.418688f);
             var bCrMult = new Vector<float>(0.081312f);
 
-            nint n = values.Component0.Length / Vector<float>.Count;
+            nint n = (nint)(uint)values.Component0.Length / Vector<float>.Count;
             for (nint i = 0; i < n; i++)
             {
                 Vector<float> r = maxSampleValue - Unsafe.Add(ref srcR, i);

src/ImageSharp/Formats/Jpeg/Components/Encoder/ComponentProcessor.cs

@@ -122,7 +122,7 @@ internal class ComponentProcessor : IDisposable
                 ref Vector256<float> sourceVectorRef = ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(source));
 
                 // Spans are guaranteed to be multiple of 8 so no extra 'remainder' steps are needed
-                nint count = source.Length / Vector256<float>.Count;
+                nint count = (nint)(uint)source.Length / Vector256<float>.Count;
                 for (nint i = 0; i < count; i++)
                 {
                     Unsafe.Add(ref targetVectorRef, i) = Avx.Add(Unsafe.Add(ref targetVectorRef, i), Unsafe.Add(ref sourceVectorRef, i));
@@ -133,7 +133,7 @@ internal class ComponentProcessor : IDisposable
                 ref Vector<float> targetVectorRef = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(target));
                 ref Vector<float> sourceVectorRef = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(source));
 
-                nint count = source.Length / Vector<float>.Count;
+                nint count = (nint)(uint)source.Length / Vector<float>.Count;
                 for (nint i = 0; i < count; i++)
                 {
                     Unsafe.Add(ref targetVectorRef, i) += Unsafe.Add(ref sourceVectorRef, i);
@@ -200,7 +200,7 @@ internal class ComponentProcessor : IDisposable
                 ref Vector256<float> targetVectorRef = ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(target));
 
                 // Spans are guaranteed to be multiple of 8 so no extra 'remainder' steps are needed
-                nint count = target.Length / Vector256<float>.Count;
+                nint count = (nint)(uint)target.Length / Vector256<float>.Count;
                 var multiplierVector = Vector256.Create(multiplier);
                 for (nint i = 0; i < count; i++)
                 {
@@ -211,7 +211,7 @@ internal class ComponentProcessor : IDisposable
             {
                 ref Vector<float> targetVectorRef = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(target));
 
-                nint count = target.Length / Vector<float>.Count;
+                nint count = (nint)(uint)target.Length / Vector<float>.Count;
                 var multiplierVector = new Vector<float>(multiplier);
                 for (nint i = 0; i < count; i++)
                 {

tests/ImageSharp.Benchmarks/Bulk/FromVector4.cs

@@ -103,7 +103,7 @@ public class FromVector4Rgba32 : FromVector4<Rgba32>
         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;
+        nint n = (nint)(uint)dest.Length / Vector<byte>.Count;
 
         ref Vector256<float> sourceBase =
             ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(src));
@@ -114,7 +114,7 @@ public class FromVector4Rgba32 : FromVector4<Rgba32>
 
         var maxBytes = Vector256.Create(255f);
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             ref Vector256<float> s = ref Unsafe.Add(ref sourceBase, i * 4);
 

tests/ImageSharp.Benchmarks/Bulk/ToVector4_Rgba32.cs

@@ -54,13 +54,13 @@ public class ToVector4_Rgba32 : ToVector4<Rgba32>
         Span<byte> sBytes = MemoryMarshal.Cast<Rgba32, byte>(this.source.GetSpan());
         Span<float> dFloats = MemoryMarshal.Cast<Vector4, float>(this.destination.GetSpan());
 
-        int n = dFloats.Length / Vector<byte>.Count;
+        nint n = (nint)(uint)dFloats.Length / Vector<byte>.Count;
 
         ref Vector<byte> sourceBase = ref Unsafe.As<byte, Vector<byte>>(ref MemoryMarshal.GetReference((ReadOnlySpan<byte>)sBytes));
         ref Vector<float> destBase = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(dFloats));
         ref Vector<uint> destBaseU = ref Unsafe.As<Vector<float>, Vector<uint>>(ref destBase);
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             Vector<byte> b = Unsafe.Add(ref sourceBase, i);
 
@@ -75,10 +75,10 @@ public class ToVector4_Rgba32 : ToVector4<Rgba32>
             Unsafe.Add(ref d, 3) = w3;
         }
 
-        n = dFloats.Length / Vector<float>.Count;
+        n = (nint)(uint)dFloats.Length / Vector<float>.Count;
         var scale = new Vector<float>(1f / 255f);
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             ref Vector<float> dRef = ref Unsafe.Add(ref destBase, i);
 
@@ -96,13 +96,13 @@ public class ToVector4_Rgba32 : ToVector4<Rgba32>
         Span<byte> sBytes = MemoryMarshal.Cast<Rgba32, byte>(this.source.GetSpan());
         Span<float> dFloats = MemoryMarshal.Cast<Vector4, float>(this.destination.GetSpan());
 
-        int n = dFloats.Length / Vector<byte>.Count;
+        nint n = (nint)(uint)dFloats.Length / Vector<byte>.Count;
 
         ref Vector<byte> sourceBase = ref Unsafe.As<byte, Vector<byte>>(ref MemoryMarshal.GetReference((ReadOnlySpan<byte>)sBytes));
         ref Vector<float> destBase = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(dFloats));
         var scale = new Vector<float>(1f / 255f);
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             Vector<byte> b = Unsafe.Add(ref sourceBase, i);
 

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

@@ -205,14 +205,14 @@ public unsafe class PixelConversion_PackFromRgbPlanes
         ref Vector256<float> bBase = ref Unsafe.As<float, Vector256<float>>(ref this.bFloat[0]);
         ref Vector256<float> resultBase = ref Unsafe.As<float, Vector256<float>>(ref this.rgbaFloat[0]);
 
-        int count = this.Count / Vector256<float>.Count;
+        nint count = (nint)(uint)this.Count / Vector256<float>.Count;
 
         ref byte control = ref MemoryMarshal.GetReference(SimdUtils.HwIntrinsics.PermuteMaskEvenOdd8x32);
         Vector256<int> vcontrol = Unsafe.As<byte, Vector256<int>>(ref control);
 
         var va = Vector256.Create(1F);
 
-        for (int i = 0; i < count; i++)
+        for (nint i = 0; i < count; i++)
         {
             Vector256<float> r = Unsafe.Add(ref rBase, i);
             Vector256<float> g = Unsafe.Add(ref gBase, i);

tests/ImageSharp.Benchmarks/General/Vectorization/UInt32ToSingle.cs

@@ -25,14 +25,14 @@ public class UInt32ToSingle
     {
         ref Vector<float> b = ref Unsafe.As<float, Vector<float>>(ref this.data[0]);
 
-        int n = Count / Vector<float>.Count;
+        nint n = Count / Vector<float>.Count;
 
         var bVec = new Vector<float>(256.0f / 255.0f);
         var magicFloat = new Vector<float>(32768.0f);
         var magicInt = new Vector<uint>(1191182336); // reinterpreted value of 32768.0f
         var mask = new Vector<uint>(255);
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             ref Vector<float> df = ref Unsafe.Add(ref b, i);
 
@@ -50,14 +50,14 @@ public class UInt32ToSingle
     [Benchmark]
     public void StandardSimd()
     {
-        int n = Count / Vector<float>.Count;
+        nint n = Count / Vector<float>.Count;
 
         ref Vector<float> bf = ref Unsafe.As<float, Vector<float>>(ref this.data[0]);
         ref Vector<uint> bu = ref Unsafe.As<Vector<float>, Vector<uint>>(ref bf);
 
         var scale = new Vector<float>(1f / 255f);
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             Vector<uint> u = Unsafe.Add(ref bu, i);
             Vector<float> v = Vector.ConvertToSingle(u);
@@ -69,14 +69,14 @@ public class UInt32ToSingle
     [Benchmark]
     public void StandardSimdFromInt()
     {
-        int n = Count / Vector<float>.Count;
+        nint n = Count / Vector<float>.Count;
 
         ref Vector<float> bf = ref Unsafe.As<float, Vector<float>>(ref this.data[0]);
         ref Vector<int> bu = ref Unsafe.As<Vector<float>, Vector<int>>(ref bf);
 
         var scale = new Vector<float>(1f / 255f);
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             Vector<int> u = Unsafe.Add(ref bu, i);
             Vector<float> v = Vector.ConvertToSingle(u);
@@ -88,12 +88,12 @@ public class UInt32ToSingle
     [Benchmark]
     public void StandardSimdFromInt_RefCast()
     {
-        int n = Count / Vector<float>.Count;
+        nint n = Count / Vector<float>.Count;
 
         ref Vector<float> bf = ref Unsafe.As<float, Vector<float>>(ref this.data[0]);
         var scale = new Vector<float>(1f / 255f);
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             ref Vector<float> fRef = ref Unsafe.Add(ref bf, i);
 

tests/ImageSharp.Benchmarks/General/Vectorization/VectorFetching.cs

@@ -63,14 +63,14 @@ public class VectorFetching
         var v = new Vector<float>(this.testValue);
         ref Vector<float> start = ref Unsafe.As<float, Vector<float>>(ref this.data[0]);
 
-        int n = this.InputSize / Vector<uint>.Count;
+        nint n = (nint)(uint)this.InputSize / Vector<uint>.Count;
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             ref Vector<float> p = ref Unsafe.Add(ref start, i);
 
             Vector<float> a = p;
-            a = a * v;
+            a *= v;
 
             p = a;
         }
@@ -82,12 +82,12 @@ public class VectorFetching
         var v = new Vector<float>(this.testValue);
         ref Vector<float> start = ref Unsafe.As<float, Vector<float>>(ref this.data[0]);
 
-        int n = this.InputSize / Vector<uint>.Count;
+        nint n = (nint)(uint)this.InputSize / Vector<uint>.Count;
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             ref Vector<float> a = ref Unsafe.Add(ref start, i);
-            a = a * v;
+            a *= v;
         }
     }
 
@@ -100,12 +100,12 @@ public class VectorFetching
 
         ref Vector<float> start = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(span));
 
-        int n = this.InputSize / Vector<uint>.Count;
+        nint n = (nint)(uint)this.InputSize / Vector<uint>.Count;
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             ref Vector<float> a = ref Unsafe.Add(ref start, i);
-            a = a * v;
+            a *= v;
         }
     }
 }

tests/ImageSharp.Benchmarks/General/Vectorization/WidenBytesToUInt32.cs

@@ -42,12 +42,12 @@ public class WidenBytesToUInt32
     [Benchmark]
     public void Simd()
     {
-        int n = Count / Vector<byte>.Count;
+        nint n = Count / Vector<byte>.Count;
 
         ref Vector<byte> sBase = ref Unsafe.As<byte, Vector<byte>>(ref this.source[0]);
         ref Vector<uint> dBase = ref Unsafe.As<uint, Vector<uint>>(ref this.dest[0]);
 
-        for (int i = 0; i < n; i++)
+        for (nint i = 0; i < n; i++)
         {
             Vector<byte> b = Unsafe.Add(ref sBase, i);