Use tensor negation for sharpen kernels

src/ImageSharp/Common/Helpers/TensorPrimitives.cs

@@ -16,7 +16,7 @@ namespace SixLabors.ImageSharp.Common.Helpers;
 /// implementation when ImageSharp no longer supports target frameworks that predate it.
 /// </remarks>
 #pragma warning disable SA1649 // File name should match first type name
-internal static class TensorPrimitives_
+internal static partial class TensorPrimitives_
 #pragma warning restore SA1649 // File name should match first type name
 {
     /// <summary>

src/ImageSharp/Common/Helpers/TensorPrimitives_.Negate.cs

@@ -0,0 +1,276 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
+
+namespace SixLabors.ImageSharp.Common.Helpers;
+
+internal static partial class TensorPrimitives_
+{
+    /// <summary>
+    /// Defines an element-wise unary operation.
+    /// </summary>
+    /// <typeparam name="T">The element type.</typeparam>
+    private interface IUnaryOperator<T>
+    {
+        /// <summary>
+        /// Gets a value indicating whether the operation supports vector execution.
+        /// </summary>
+        public static abstract bool Vectorizable { get; }
+
+        /// <summary>
+        /// Applies the operation to a scalar value.
+        /// </summary>
+        /// <param name="x">The input value.</param>
+        /// <returns>The operation result.</returns>
+        public static abstract T Invoke(T x);
+
+        /// <summary>
+        /// Applies the operation to a 128-bit vector.
+        /// </summary>
+        /// <param name="x">The input vector.</param>
+        /// <returns>The operation result.</returns>
+        public static abstract Vector128<T> Invoke(Vector128<T> x);
+
+        /// <summary>
+        /// Applies the operation to a 256-bit vector.
+        /// </summary>
+        /// <param name="x">The input vector.</param>
+        /// <returns>The operation result.</returns>
+        public static abstract Vector256<T> Invoke(Vector256<T> x);
+
+        /// <summary>
+        /// Applies the operation to a 512-bit vector.
+        /// </summary>
+        /// <param name="x">The input vector.</param>
+        /// <returns>The operation result.</returns>
+        public static abstract Vector512<T> Invoke(Vector512<T> x);
+    }
+
+    /// <summary>
+    /// Computes the element-wise negation of the values in <paramref name="x"/>.
+    /// </summary>
+    /// <typeparam name="T">The element type.</typeparam>
+    /// <param name="x">The values to negate.</param>
+    /// <param name="destination">The destination for the negated values.</param>
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static void Negate<T>(ReadOnlySpan<T> x, Span<T> destination)
+        where T : IUnaryNegationOperators<T, T>
+        => InvokeSpanIntoSpan<T, NegateOperator<T>>(x, destination);
+
+    /// <summary>
+    /// Performs an element-wise unary operation over a span.
+    /// </summary>
+    /// <typeparam name="T">The element type.</typeparam>
+    /// <typeparam name="TOperator">The operation to apply.</typeparam>
+    /// <param name="x">The input values.</param>
+    /// <param name="destination">The destination values.</param>
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static void InvokeSpanIntoSpan<T, TOperator>(ReadOnlySpan<T> x, Span<T> destination)
+        where TOperator : struct, IUnaryOperator<T>
+    {
+        ref T xRef = ref MemoryMarshal.GetReference(x);
+        ref T destinationRef = ref MemoryMarshal.GetReference(destination);
+        nuint length = (uint)x.Length;
+
+        // The dispatch matches the other compatibility pipelines: AVX-512 is reserved for large spans because
+        // its setup cost is not recovered by the short image-processing buffers that dominate ImageSharp.
+        if (TOperator.Vectorizable
+            && Vector512.IsHardwareAccelerated
+            && Vector512<T>.IsSupported
+            && length >= 512)
+        {
+            InvokeUnaryVectorized512<T, TOperator>(ref xRef, ref destinationRef, length);
+            return;
+        }
+
+        if (TOperator.Vectorizable && Vector256.IsHardwareAccelerated && Vector256<T>.IsSupported && length >= (uint)Vector256<T>.Count)
+        {
+            InvokeUnaryVectorized256<T, TOperator>(ref xRef, ref destinationRef, length);
+            return;
+        }
+
+        if (TOperator.Vectorizable && Vector128.IsHardwareAccelerated && Vector128<T>.IsSupported && length >= (uint)Vector128<T>.Count)
+        {
+            InvokeUnaryVectorized128<T, TOperator>(ref xRef, ref destinationRef, length);
+            return;
+        }
+
+        for (nuint i = 0; i < length; i++)
+        {
+            Unsafe.Add(ref destinationRef, i) = TOperator.Invoke(Unsafe.Add(ref xRef, i));
+        }
+    }
+
+    /// <summary>
+    /// Applies a unary operation with 128-bit vectors.
+    /// </summary>
+    /// <typeparam name="T">The element type.</typeparam>
+    /// <typeparam name="TOperator">The operation to apply.</typeparam>
+    /// <param name="xRef">The first input element.</param>
+    /// <param name="destinationRef">The first destination element.</param>
+    /// <param name="length">The number of elements to process.</param>
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static void InvokeUnaryVectorized128<T, TOperator>(ref T xRef, ref T destinationRef, nuint length)
+        where TOperator : struct, IUnaryOperator<T>
+    {
+        nuint vectorCount = (uint)Vector128<T>.Count;
+        nuint vectorsPerLoop = vectorCount * 8;
+        nuint index = 0;
+
+        // The final vector overlaps the preceding store when the length is not a vector multiple. Loading it
+        // before any stores preserves same-start in-place operation because it captures the original tail.
+        Vector128<T> end = default;
+        if ((length % vectorCount) != 0)
+        {
+            end = TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, length - vectorCount));
+        }
+
+        while ((length - index) >= vectorsPerLoop)
+        {
+            TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, index + (vectorCount * 0))).StoreUnsafe(ref destinationRef, index + (vectorCount * 0));
+            TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, index + (vectorCount * 1))).StoreUnsafe(ref destinationRef, index + (vectorCount * 1));
+            TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, index + (vectorCount * 2))).StoreUnsafe(ref destinationRef, index + (vectorCount * 2));
+            TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, index + (vectorCount * 3))).StoreUnsafe(ref destinationRef, index + (vectorCount * 3));
+            TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, index + (vectorCount * 4))).StoreUnsafe(ref destinationRef, index + (vectorCount * 4));
+            TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, index + (vectorCount * 5))).StoreUnsafe(ref destinationRef, index + (vectorCount * 5));
+            TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, index + (vectorCount * 6))).StoreUnsafe(ref destinationRef, index + (vectorCount * 6));
+            TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, index + (vectorCount * 7))).StoreUnsafe(ref destinationRef, index + (vectorCount * 7));
+
+            index += vectorsPerLoop;
+        }
+
+        while ((length - index) >= vectorCount)
+        {
+            TOperator.Invoke(Vector128.LoadUnsafe(ref xRef, index)).StoreUnsafe(ref destinationRef, index);
+            index += vectorCount;
+        }
+
+        if (index != length)
+        {
+            end.StoreUnsafe(ref destinationRef, length - vectorCount);
+        }
+    }
+
+    /// <summary>
+    /// Applies a unary operation with 256-bit vectors.
+    /// </summary>
+    /// <typeparam name="T">The element type.</typeparam>
+    /// <typeparam name="TOperator">The operation to apply.</typeparam>
+    /// <param name="xRef">The first input element.</param>
+    /// <param name="destinationRef">The first destination element.</param>
+    /// <param name="length">The number of elements to process.</param>
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static void InvokeUnaryVectorized256<T, TOperator>(ref T xRef, ref T destinationRef, nuint length)
+        where TOperator : struct, IUnaryOperator<T>
+    {
+        nuint vectorCount = (uint)Vector256<T>.Count;
+        nuint vectorsPerLoop = vectorCount * 8;
+        nuint index = 0;
+        Vector256<T> end = default;
+
+        if ((length % vectorCount) != 0)
+        {
+            end = TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, length - vectorCount));
+        }
+
+        while ((length - index) >= vectorsPerLoop)
+        {
+            TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, index + (vectorCount * 0))).StoreUnsafe(ref destinationRef, index + (vectorCount * 0));
+            TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, index + (vectorCount * 1))).StoreUnsafe(ref destinationRef, index + (vectorCount * 1));
+            TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, index + (vectorCount * 2))).StoreUnsafe(ref destinationRef, index + (vectorCount * 2));
+            TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, index + (vectorCount * 3))).StoreUnsafe(ref destinationRef, index + (vectorCount * 3));
+            TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, index + (vectorCount * 4))).StoreUnsafe(ref destinationRef, index + (vectorCount * 4));
+            TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, index + (vectorCount * 5))).StoreUnsafe(ref destinationRef, index + (vectorCount * 5));
+            TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, index + (vectorCount * 6))).StoreUnsafe(ref destinationRef, index + (vectorCount * 6));
+            TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, index + (vectorCount * 7))).StoreUnsafe(ref destinationRef, index + (vectorCount * 7));
+
+            index += vectorsPerLoop;
+        }
+
+        while ((length - index) >= vectorCount)
+        {
+            TOperator.Invoke(Vector256.LoadUnsafe(ref xRef, index)).StoreUnsafe(ref destinationRef, index);
+            index += vectorCount;
+        }
+
+        if (index != length)
+        {
+            end.StoreUnsafe(ref destinationRef, length - vectorCount);
+        }
+    }
+
+    /// <summary>
+    /// Applies a unary operation with 512-bit vectors.
+    /// </summary>
+    /// <typeparam name="T">The element type.</typeparam>
+    /// <typeparam name="TOperator">The operation to apply.</typeparam>
+    /// <param name="xRef">The first input element.</param>
+    /// <param name="destinationRef">The first destination element.</param>
+    /// <param name="length">The number of elements to process.</param>
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static void InvokeUnaryVectorized512<T, TOperator>(ref T xRef, ref T destinationRef, nuint length)
+        where TOperator : struct, IUnaryOperator<T>
+    {
+        nuint vectorCount = (uint)Vector512<T>.Count;
+        nuint vectorsPerLoop = vectorCount * 8;
+        nuint index = 0;
+        Vector512<T> end = default;
+
+        if ((length % vectorCount) != 0)
+        {
+            end = TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, length - vectorCount));
+        }
+
+        while ((length - index) >= vectorsPerLoop)
+        {
+            TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, index + (vectorCount * 0))).StoreUnsafe(ref destinationRef, index + (vectorCount * 0));
+            TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, index + (vectorCount * 1))).StoreUnsafe(ref destinationRef, index + (vectorCount * 1));
+            TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, index + (vectorCount * 2))).StoreUnsafe(ref destinationRef, index + (vectorCount * 2));
+            TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, index + (vectorCount * 3))).StoreUnsafe(ref destinationRef, index + (vectorCount * 3));
+            TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, index + (vectorCount * 4))).StoreUnsafe(ref destinationRef, index + (vectorCount * 4));
+            TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, index + (vectorCount * 5))).StoreUnsafe(ref destinationRef, index + (vectorCount * 5));
+            TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, index + (vectorCount * 6))).StoreUnsafe(ref destinationRef, index + (vectorCount * 6));
+            TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, index + (vectorCount * 7))).StoreUnsafe(ref destinationRef, index + (vectorCount * 7));
+
+            index += vectorsPerLoop;
+        }
+
+        while ((length - index) >= vectorCount)
+        {
+            TOperator.Invoke(Vector512.LoadUnsafe(ref xRef, index)).StoreUnsafe(ref destinationRef, index);
+            index += vectorCount;
+        }
+
+        if (index != length)
+        {
+            end.StoreUnsafe(ref destinationRef, length - vectorCount);
+        }
+    }
+
+    /// <summary>
+    /// Implements element-wise negation for scalar and SIMD inputs.
+    /// </summary>
+    /// <typeparam name="T">The element type.</typeparam>
+    private readonly struct NegateOperator<T> : IUnaryOperator<T>
+        where T : IUnaryNegationOperators<T, T>
+    {
+        /// <inheritdoc />
+        public static bool Vectorizable => true;
+
+        /// <inheritdoc />
+        public static T Invoke(T x) => -x;
+
+        /// <inheritdoc />
+        public static Vector128<T> Invoke(Vector128<T> x) => -x;
+
+        /// <inheritdoc />
+        public static Vector256<T> Invoke(Vector256<T> x) => -x;
+
+        /// <inheritdoc />
+        public static Vector512<T> Invoke(Vector512<T> x) => -x;
+    }
+}

src/ImageSharp/Processing/Processors/Convolution/ConvolutionProcessorHelpers.cs

@@ -2,6 +2,7 @@
 // Licensed under the Six Labors Split License.
 
 using System.Diagnostics.CodeAnalysis;
+using SixLabors.ImageSharp.Common.Helpers;
 
 namespace SixLabors.ImageSharp.Processing.Processors.Convolution;
 
@@ -68,19 +69,12 @@ internal static class ConvolutionProcessorHelpers
 
         // Invert the kernel for sharpening.
         int midpointRounded = (int)midpoint;
-        for (int i = 0; i < size; i++)
+        float midpointValue = kernel[midpointRounded];
-        {
+        TensorPrimitives_.Negate<float>(kernel, kernel);
-            if (i == midpointRounded)
+
-            {
+        // The sharpening kernel negates every Gaussian weight except its center. Restore that original
-                // Calculate central value
+        // center while adding twice the Gaussian sum so the complete kernel retains unit response.
-                kernel[i] = (2F * sum) - kernel[i];
+        kernel[midpointRounded] = (2F * sum) - midpointValue;
-            }
-            else
-            {
-                // invert value
-                kernel[i] = -kernel[i];
-            }
-        }
 
         // Normalize kernel so that the sum of all weights equals 1
         for (int i = 0; i < size; i++)

tests/ImageSharp.Tests/Common/TensorPrimitivesTests.cs

@@ -117,6 +117,42 @@ public class TensorPrimitivesTests
         Assert.Equal(expected, inPlace);
     }
 
+    /// <summary>
+    /// Verifies that floating-point negation preserves the scalar operator's exact bit-level behavior.
+    /// </summary>
+    /// <param name="length">The input length.</param>
+    [Theory]
+    [MemberData(nameof(SpanLengths))]
+    public void NegateSingleMatchesScalarFormula(int length)
+    {
+        float[] values =
+        {
+            float.NaN,
+            -0F,
+            0F,
+            -1F,
+            1F,
+            float.NegativeInfinity,
+            float.PositiveInfinity
+        };
+
+        float[] source = new float[length];
+        float[] expected = new float[length];
+
+        for (int i = 0; i < source.Length; i++)
+        {
+            source[i] = values[i % values.Length];
+            expected[i] = -source[i];
+        }
+
+        float[] destination = new float[length];
+        TensorPrimitives_.Negate<float>(source, destination);
+        AssertSingleBitsEqual(expected, destination);
+
+        TensorPrimitives_.Negate<float>(source, source);
+        AssertSingleBitsEqual(expected, source);
+    }
+
     /// <summary>
     /// Verifies that integer clamping produces identical results for separate and in-place destinations.
     /// </summary>

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

@@ -41,6 +41,41 @@ public class ConvolutionProcessorHelpersTest
         }
     }
 
+    /// <summary>
+    /// Verifies that Gaussian sharpening preserves the scalar kernel formula across scalar and SIMD lengths.
+    /// </summary>
+    /// <param name="radius">The kernel radius.</param>
+    [Theory]
+    [InlineData(1)]
+    [InlineData(3)]
+    [InlineData(9)]
+    [InlineData(32)]
+    [InlineData(80)]
+    public void VerifyGaussianSharpenKernel(int radius)
+    {
+        int kernelSize = (radius * 2) + 1;
+        float sigma = radius / 3F;
+        float[] expected = new float[kernelSize];
+        float sum = 0F;
+
+        for (int i = 0; i < kernelSize; i++)
+        {
+            float value = Numerics.Gaussian(i - radius, sigma);
+            expected[i] = value;
+            sum += value;
+        }
+
+        for (int i = 0; i < kernelSize; i++)
+        {
+            expected[i] = i == radius ? (2F * sum) - expected[i] : -expected[i];
+            expected[i] /= sum;
+        }
+
+        float[] actual = ConvolutionProcessorHelpers.CreateGaussianSharpenKernel(kernelSize, sigma);
+
+        Assert.Equal(expected, actual);
+    }
+
     [Fact]
     public void VerifyNonSeparableMatrix()
     {