diff --git a/src/ImageSharp.Drawing/Processing/PatternBrush{TPixel}.cs b/src/ImageSharp.Drawing/Processing/PatternBrush{TPixel}.cs
index 46ed36f68..20161b517 100644
--- a/src/ImageSharp.Drawing/Processing/PatternBrush{TPixel}.cs
+++ b/src/ImageSharp.Drawing/Processing/PatternBrush{TPixel}.cs
@@ -61,7 +61,7 @@ namespace SixLabors.ImageSharp.Processing
         /// <param name="foreColor">Color of the fore.</param>
         /// <param name="backColor">Color of the back.</param>
         /// <param name="pattern">The pattern.</param>
-        internal PatternBrush(TPixel foreColor, TPixel backColor, DenseMatrix<bool> pattern)
+        internal PatternBrush(TPixel foreColor, TPixel backColor, in DenseMatrix<bool> pattern)
         {
             var foreColorVector = foreColor.ToVector4();
             var backColorVector = backColor.ToVector4();
@@ -93,10 +93,7 @@ namespace SixLabors.ImageSharp.Processing
         }
 
         /// <inheritdoc />
-        public BrushApplicator<TPixel> CreateApplicator(ImageFrame<TPixel> source, RectangleF region, GraphicsOptions options)
-        {
-            return new PatternBrushApplicator(source, this.pattern, this.patternVector, options);
-        }
+        public BrushApplicator<TPixel> CreateApplicator(ImageFrame<TPixel> source, RectangleF region, GraphicsOptions options) => new PatternBrushApplicator(source, this.pattern, this.patternVector, options);
 
         /// <summary>
         /// The pattern brush applicator.
@@ -116,7 +113,7 @@ namespace SixLabors.ImageSharp.Processing
             /// <param name="pattern">The pattern.</param>
             /// <param name="patternVector">The patternVector.</param>
             /// <param name="options">The options</param>
-            public PatternBrushApplicator(ImageFrame<TPixel> source, DenseMatrix<TPixel> pattern, DenseMatrix<Vector4> patternVector, GraphicsOptions options)
+            public PatternBrushApplicator(ImageFrame<TPixel> source, in DenseMatrix<TPixel> pattern, DenseMatrix<Vector4> patternVector, GraphicsOptions options)
                 : base(source, options)
             {
                 this.pattern = pattern;
diff --git a/src/ImageSharp/Common/Helpers/DenseMatrixUtils.cs b/src/ImageSharp/Common/Helpers/DenseMatrixUtils.cs
index 2e700c9d6..427b24005 100644
--- a/src/ImageSharp/Common/Helpers/DenseMatrixUtils.cs
+++ b/src/ImageSharp/Common/Helpers/DenseMatrixUtils.cs
@@ -3,6 +3,7 @@
 
 using System;
 using System.Numerics;
+using System.Runtime.CompilerServices;
 using SixLabors.ImageSharp.Memory;
 using SixLabors.ImageSharp.PixelFormats;
 using SixLabors.ImageSharp.Primitives;
@@ -11,83 +12,120 @@ namespace SixLabors.ImageSharp
 {
     /// <summary>
     /// Extension methods for <see cref="DenseMatrix{T}"/>.
+    /// TODO: One day rewrite all this to use SIMD intrinsics. There's a lot of scope for improvement.
     /// </summary>
     internal static class DenseMatrixUtils
     {
         /// <summary>
-        /// Computes the sum of vectors in <paramref name="targetRow"/> weighted by the kernel weight values.
+        /// Computes the sum of vectors in the span referenced by <paramref name="targetRowRef"/> weighted by the two kernel weight values.
+        /// Using this method the convolution filter is not applied to alpha in addition to the color channels.
         /// </summary>
         /// <typeparam name="TPixel">The pixel format.</typeparam>
-        /// <param name="matrix">The dense matrix.</param>
+        /// <param name="matrixY">The vertical dense matrix.</param>
+        /// <param name="matrixX">The horizontal dense matrix.</param>
         /// <param name="sourcePixels">The source frame.</param>
-        /// <param name="targetRow">The target row.</param>
+        /// <param name="targetRowRef">The target row base reference.</param>
         /// <param name="row">The current row.</param>
         /// <param name="column">The current column.</param>
+        /// <param name="minRow">The minimum working area row.</param>
         /// <param name="maxRow">The maximum working area row.</param>
+        /// <param name="minColumn">The minimum working area column.</param>
         /// <param name="maxColumn">The maximum working area column.</param>
-        /// <param name="offsetColumn">The column offset to apply to source sampling.</param>
-        public static void Convolve<TPixel>(
-            in DenseMatrix<float> matrix,
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public static void Convolve2D3<TPixel>(
+            in DenseMatrix<float> matrixY,
+            in DenseMatrix<float> matrixX,
             Buffer2D<TPixel> sourcePixels,
-            Span<Vector4> targetRow,
+            ref Vector4 targetRowRef,
             int row,
             int column,
+            int minRow,
             int maxRow,
-            int maxColumn,
-            int offsetColumn)
+            int minColumn,
+            int maxColumn)
             where TPixel : struct, IPixel<TPixel>
         {
             Vector4 vector = default;
-            int matrixHeight = matrix.Rows;
-            int matrixWidth = matrix.Columns;
-            int radiusY = matrixHeight >> 1;
-            int radiusX = matrixWidth >> 1;
-            int sourceOffsetColumnBase = column + offsetColumn;
-
-            for (int y = 0; y < matrixHeight; y++)
-            {
-                int offsetY = (row + y - radiusY).Clamp(0, maxRow);
-                Span<TPixel> sourceRowSpan = sourcePixels.GetRowSpan(offsetY);
 
-                for (int x = 0; x < matrixWidth; x++)
-                {
-                    int offsetX = (sourceOffsetColumnBase + x - radiusX).Clamp(offsetColumn, maxColumn);
-                    var currentColor = sourceRowSpan[offsetX].ToVector4();
-                    Vector4Utils.Premultiply(ref currentColor);
-
-                    vector += matrix[y, x] * currentColor;
-                }
-            }
+            Convolve2DImpl(
+                in matrixY,
+                in matrixX,
+                sourcePixels,
+                row,
+                column,
+                minRow,
+                maxRow,
+                minColumn,
+                maxColumn,
+                ref vector);
 
-            ref Vector4 target = ref targetRow[column];
+            ref Vector4 target = ref Unsafe.Add(ref targetRowRef, column);
             vector.W = target.W;
+
             Vector4Utils.UnPremultiply(ref vector);
             target = vector;
         }
 
         /// <summary>
-        /// Computes the sum of vectors in <paramref name="targetRow"/> weighted by the two kernel weight values.
+        /// Computes the sum of vectors in the span referenced by <paramref name="targetRowRef"/> weighted by the two kernel weight values.
+        /// Using this method the convolution filter is applied to alpha in addition to the color channels.
         /// </summary>
         /// <typeparam name="TPixel">The pixel format.</typeparam>
         /// <param name="matrixY">The vertical dense matrix.</param>
         /// <param name="matrixX">The horizontal dense matrix.</param>
         /// <param name="sourcePixels">The source frame.</param>
-        /// <param name="targetRow">The target row.</param>
+        /// <param name="targetRowRef">The target row base reference.</param>
         /// <param name="row">The current row.</param>
         /// <param name="column">The current column.</param>
+        /// <param name="minRow">The minimum working area row.</param>
         /// <param name="maxRow">The maximum working area row.</param>
+        /// <param name="minColumn">The minimum working area column.</param>
         /// <param name="maxColumn">The maximum working area column.</param>
-        /// <param name="offsetColumn">The column offset to apply to source sampling.</param>
-        public static void Convolve2D<TPixel>(
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public static void Convolve2D4<TPixel>(
             in DenseMatrix<float> matrixY,
             in DenseMatrix<float> matrixX,
             Buffer2D<TPixel> sourcePixels,
-            Span<Vector4> targetRow,
+            ref Vector4 targetRowRef,
             int row,
             int column,
+            int minRow,
             int maxRow,
+            int minColumn,
+            int maxColumn)
+            where TPixel : struct, IPixel<TPixel>
+        {
+            Vector4 vector = default;
+
+            Convolve2DImpl(
+                in matrixY,
+                in matrixX,
+                sourcePixels,
+                row,
+                column,
+                minRow,
+                maxRow,
+                minColumn,
+                maxColumn,
+                ref vector);
+
+            ref Vector4 target = ref Unsafe.Add(ref targetRowRef, column);
+            Vector4Utils.UnPremultiply(ref vector);
+            target = vector;
+        }
+
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public static void Convolve2DImpl<TPixel>(
+            in DenseMatrix<float> matrixY,
+            in DenseMatrix<float> matrixX,
+            Buffer2D<TPixel> sourcePixels,
+            int row,
+            int column,
+            int minRow,
+            int maxRow,
+            int minColumn,
             int maxColumn,
-            int offsetColumn)
+            ref Vector4 vector)
             where TPixel : struct, IPixel<TPixel>
         {
             Vector4 vectorY = default;
@@ -96,16 +134,16 @@ namespace SixLabors.ImageSharp
             int matrixWidth = matrixY.Columns;
             int radiusY = matrixHeight >> 1;
             int radiusX = matrixWidth >> 1;
-            int sourceOffsetColumnBase = column + offsetColumn;
+            int sourceOffsetColumnBase = column + minColumn;
 
             for (int y = 0; y < matrixHeight; y++)
             {
-                int offsetY = (row + y - radiusY).Clamp(0, maxRow);
+                int offsetY = (row + y - radiusY).Clamp(minRow, maxRow);
                 Span<TPixel> sourceRowSpan = sourcePixels.GetRowSpan(offsetY);
 
                 for (int x = 0; x < matrixWidth; x++)
                 {
-                    int offsetX = (sourceOffsetColumnBase + x - radiusX).Clamp(offsetColumn, maxColumn);
+                    int offsetX = (sourceOffsetColumnBase + x - radiusX).Clamp(minColumn, maxColumn);
                     var currentColor = sourceRowSpan[offsetX].ToVector4();
                     Vector4Utils.Premultiply(ref currentColor);
 
@@ -114,11 +152,133 @@ namespace SixLabors.ImageSharp
                 }
             }
 
-            var vector = Vector4.SquareRoot((vectorX * vectorX) + (vectorY * vectorY));
-            ref Vector4 target = ref targetRow[column];
+            vector = Vector4.SquareRoot((vectorX * vectorX) + (vectorY * vectorY));
+        }
+
+        /// <summary>
+        /// Computes the sum of vectors in the span referenced by <paramref name="targetRowRef"/> weighted by the kernel weight values.
+        /// Using this method the convolution filter is not applied to alpha in addition to the color channels.
+        /// </summary>
+        /// <typeparam name="TPixel">The pixel format.</typeparam>
+        /// <param name="matrix">The dense matrix.</param>
+        /// <param name="sourcePixels">The source frame.</param>
+        /// <param name="targetRowRef">The target row base reference.</param>
+        /// <param name="row">The current row.</param>
+        /// <param name="column">The current column.</param>
+        /// <param name="minRow">The minimum working area row.</param>
+        /// <param name="maxRow">The maximum working area row.</param>
+        /// <param name="minColumn">The minimum working area column.</param>
+        /// <param name="maxColumn">The maximum working area column.</param>
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public static void Convolve3<TPixel>(
+            in DenseMatrix<float> matrix,
+            Buffer2D<TPixel> sourcePixels,
+            ref Vector4 targetRowRef,
+            int row,
+            int column,
+            int minRow,
+            int maxRow,
+            int minColumn,
+            int maxColumn)
+            where TPixel : struct, IPixel<TPixel>
+        {
+            Vector4 vector = default;
+
+            ConvolveImpl(
+                in matrix,
+                sourcePixels,
+                row,
+                column,
+                minRow,
+                maxRow,
+                minColumn,
+                maxColumn,
+                ref vector);
+
+            ref Vector4 target = ref Unsafe.Add(ref targetRowRef, column);
             vector.W = target.W;
+
             Vector4Utils.UnPremultiply(ref vector);
             target = vector;
         }
+
+        /// <summary>
+        /// Computes the sum of vectors in the span referenced by <paramref name="targetRowRef"/> weighted by the kernel weight values.
+        /// Using this method the convolution filter is applied to alpha in addition to the color channels.
+        /// </summary>
+        /// <typeparam name="TPixel">The pixel format.</typeparam>
+        /// <param name="matrix">The dense matrix.</param>
+        /// <param name="sourcePixels">The source frame.</param>
+        /// <param name="targetRowRef">The target row base reference.</param>
+        /// <param name="row">The current row.</param>
+        /// <param name="column">The current column.</param>
+        /// <param name="minRow">The minimum working area row.</param>
+        /// <param name="maxRow">The maximum working area row.</param>
+        /// <param name="minColumn">The minimum working area column.</param>
+        /// <param name="maxColumn">The maximum working area column.</param>
+        [MethodImpl(InliningOptions.ShortMethod)]
+        public static void Convolve4<TPixel>(
+            in DenseMatrix<float> matrix,
+            Buffer2D<TPixel> sourcePixels,
+            ref Vector4 targetRowRef,
+            int row,
+            int column,
+            int minRow,
+            int maxRow,
+            int minColumn,
+            int maxColumn)
+            where TPixel : struct, IPixel<TPixel>
+        {
+            Vector4 vector = default;
+
+            ConvolveImpl(
+                in matrix,
+                sourcePixels,
+                row,
+                column,
+                minRow,
+                maxRow,
+                minColumn,
+                maxColumn,
+                ref vector);
+
+            ref Vector4 target = ref Unsafe.Add(ref targetRowRef, column);
+            Vector4Utils.UnPremultiply(ref vector);
+            target = vector;
+        }
+
+        [MethodImpl(InliningOptions.ShortMethod)]
+        private static void ConvolveImpl<TPixel>(
+            in DenseMatrix<float> matrix,
+            Buffer2D<TPixel> sourcePixels,
+            int row,
+            int column,
+            int minRow,
+            int maxRow,
+            int minColumn,
+            int maxColumn,
+            ref Vector4 vector)
+            where TPixel : struct, IPixel<TPixel>
+        {
+            int matrixHeight = matrix.Rows;
+            int matrixWidth = matrix.Columns;
+            int radiusY = matrixHeight >> 1;
+            int radiusX = matrixWidth >> 1;
+            int sourceOffsetColumnBase = column + minColumn;
+
+            for (int y = 0; y < matrixHeight; y++)
+            {
+                int offsetY = (row + y - radiusY).Clamp(minRow, maxRow);
+                Span<TPixel> sourceRowSpan = sourcePixels.GetRowSpan(offsetY);
+
+                for (int x = 0; x < matrixWidth; x++)
+                {
+                    int offsetX = (sourceOffsetColumnBase + x - radiusX).Clamp(minColumn, maxColumn);
+                    var currentColor = sourceRowSpan[offsetX].ToVector4();
+                    Vector4Utils.Premultiply(ref currentColor);
+                    vector += matrix[y, x] * currentColor;
+                }
+            }
+        }
     }
 }
\ No newline at end of file
diff --git a/src/ImageSharp/Processing/Processors/Convolution/BoxBlurProcessor.cs b/src/ImageSharp/Processing/Processors/Convolution/BoxBlurProcessor.cs
index 644d6c9e1..3d5bdc42a 100644
--- a/src/ImageSharp/Processing/Processors/Convolution/BoxBlurProcessor.cs
+++ b/src/ImageSharp/Processing/Processors/Convolution/BoxBlurProcessor.cs
@@ -49,7 +49,8 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         public DenseMatrix<float> KernelY { get; }
 
         /// <inheritdoc/>
-        protected override void OnFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration) => new Convolution2PassProcessor<TPixel>(this.KernelX, this.KernelY).Apply(source, sourceRectangle, configuration);
+        protected override void OnFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration)
+            => new Convolution2PassProcessor<TPixel>(this.KernelX, this.KernelY, false).Apply(source, sourceRectangle, configuration);
 
         /// <summary>
         /// Create a 1 dimensional Box kernel.
diff --git a/src/ImageSharp/Processing/Processors/Convolution/Convolution2DProcessor.cs b/src/ImageSharp/Processing/Processors/Convolution/Convolution2DProcessor.cs
index 341a26ae8..633b50a9b 100644
--- a/src/ImageSharp/Processing/Processors/Convolution/Convolution2DProcessor.cs
+++ b/src/ImageSharp/Processing/Processors/Convolution/Convolution2DProcessor.cs
@@ -3,6 +3,7 @@
 
 using System;
 using System.Numerics;
+using System.Runtime.InteropServices;
 using SixLabors.ImageSharp.Memory;
 using SixLabors.ImageSharp.ParallelUtils;
 using SixLabors.ImageSharp.PixelFormats;
@@ -23,11 +24,13 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         /// </summary>
         /// <param name="kernelX">The horizontal gradient operator.</param>
         /// <param name="kernelY">The vertical gradient operator.</param>
-        public Convolution2DProcessor(DenseMatrix<float> kernelX, DenseMatrix<float> kernelY)
+        /// <param name="preserveAlpha">Whether the convolution filter is applied to alpha as well as the color channels.</param>
+        public Convolution2DProcessor(in DenseMatrix<float> kernelX, in DenseMatrix<float> kernelY, bool preserveAlpha)
         {
             Guard.IsTrue(kernelX.Size.Equals(kernelY.Size), $"{nameof(kernelX)} {nameof(kernelY)}", "Kernel sizes must be the same.");
             this.KernelX = kernelX;
             this.KernelY = kernelY;
+            this.PreserveAlpha = preserveAlpha;
         }
 
         /// <summary>
@@ -40,6 +43,11 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         /// </summary>
         public DenseMatrix<float> KernelY { get; }
 
+        /// <summary>
+        /// Gets a value indicating whether the convolution filter is applied to alpha as well as the color channels.
+        /// </summary>
+        public bool PreserveAlpha { get; }
+
         /// <inheritdoc/>
         protected override void OnFrameApply(
             ImageFrame<TPixel> source,
@@ -48,6 +56,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         {
             DenseMatrix<float> matrixY = this.KernelY;
             DenseMatrix<float> matrixX = this.KernelX;
+            bool preserveAlpha = this.PreserveAlpha;
 
             var interest = Rectangle.Intersect(sourceRectangle, source.Bounds());
             int startY = interest.Y;
@@ -71,18 +80,49 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
                         {
                             Span<Vector4> vectorSpan = vectorBuffer.Span;
                             int length = vectorSpan.Length;
+                            ref Vector4 vectorSpanRef = ref MemoryMarshal.GetReference(vectorSpan);
 
                             for (int y = rows.Min; y < rows.Max; y++)
                             {
                                 Span<TPixel> targetRowSpan = targetPixels.GetRowSpan(y).Slice(startX);
                                 PixelOperations<TPixel>.Instance.ToVector4(configuration, targetRowSpan.Slice(0, length), vectorSpan);
 
-                                for (int x = 0; x < width; x++)
+                                if (preserveAlpha)
+                                {
+                                    for (int x = 0; x < width; x++)
+                                    {
+                                        DenseMatrixUtils.Convolve2D3(
+                                            in matrixY,
+                                            in matrixX,
+                                            source.PixelBuffer,
+                                            ref vectorSpanRef,
+                                            y,
+                                            x,
+                                            startY,
+                                            maxY,
+                                            startX,
+                                            maxX);
+                                    }
+                                }
+                                else
                                 {
-                                    DenseMatrixUtils.Convolve2D(in matrixY, in matrixX, source.PixelBuffer, vectorSpan, y, x, maxY, maxX, startX);
+                                    for (int x = 0; x < width; x++)
+                                    {
+                                        DenseMatrixUtils.Convolve2D4(
+                                            in matrixY,
+                                            in matrixX,
+                                            source.PixelBuffer,
+                                            ref vectorSpanRef,
+                                            y,
+                                            x,
+                                            startY,
+                                            maxY,
+                                            startX,
+                                            maxX);
+                                    }
                                 }
 
-                                PixelOperations<TPixel>.Instance.FromVector4Destructive(configuration, vectorSpan.Slice(0, length), targetRowSpan);
+                                PixelOperations<TPixel>.Instance.FromVector4Destructive(configuration, vectorSpan, targetRowSpan);
                             }
                         });
 
diff --git a/src/ImageSharp/Processing/Processors/Convolution/Convolution2PassProcessor.cs b/src/ImageSharp/Processing/Processors/Convolution/Convolution2PassProcessor.cs
index 9de467328..03268c9dd 100644
--- a/src/ImageSharp/Processing/Processors/Convolution/Convolution2PassProcessor.cs
+++ b/src/ImageSharp/Processing/Processors/Convolution/Convolution2PassProcessor.cs
@@ -3,7 +3,7 @@
 
 using System;
 using System.Numerics;
-
+using System.Runtime.InteropServices;
 using SixLabors.ImageSharp.Memory;
 using SixLabors.ImageSharp.ParallelUtils;
 using SixLabors.ImageSharp.PixelFormats;
@@ -24,10 +24,15 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         /// </summary>
         /// <param name="kernelX">The horizontal gradient operator.</param>
         /// <param name="kernelY">The vertical gradient operator.</param>
-        public Convolution2PassProcessor(DenseMatrix<float> kernelX, DenseMatrix<float> kernelY)
+        /// <param name="preserveAlpha">Whether the convolution filter is applied to alpha as well as the color channels.</param>
+        public Convolution2PassProcessor(
+            in DenseMatrix<float> kernelX,
+            in DenseMatrix<float> kernelY,
+            bool preserveAlpha)
         {
             this.KernelX = kernelX;
             this.KernelY = kernelY;
+            this.PreserveAlpha = preserveAlpha;
         }
 
         /// <summary>
@@ -40,13 +45,16 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         /// </summary>
         public DenseMatrix<float> KernelY { get; }
 
+        /// <summary>
+        /// Gets a value indicating whether the convolution filter is applied to alpha as well as the color channels.
+        /// </summary>
+        public bool PreserveAlpha { get; }
+
         /// <inheritdoc/>
         protected override void OnFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration)
         {
             using (Buffer2D<TPixel> firstPassPixels = configuration.MemoryAllocator.Allocate2D<TPixel>(source.Size()))
             {
-                source.CopyTo(firstPassPixels);
-
                 var interest = Rectangle.Intersect(sourceRectangle, source.Bounds());
                 this.ApplyConvolution(firstPassPixels, source.PixelBuffer, interest, this.KernelX, configuration);
                 this.ApplyConvolution(source.PixelBuffer, firstPassPixels, interest, this.KernelY, configuration);
@@ -72,6 +80,8 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
             Configuration configuration)
         {
             DenseMatrix<float> matrix = kernel;
+            bool preserveAlpha = this.PreserveAlpha;
+
             int startY = sourceRectangle.Y;
             int endY = sourceRectangle.Bottom;
             int startX = sourceRectangle.X;
@@ -89,18 +99,47 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
                     {
                         Span<Vector4> vectorSpan = vectorBuffer.Span;
                         int length = vectorSpan.Length;
+                        ref Vector4 vectorSpanRef = ref MemoryMarshal.GetReference(vectorSpan);
 
                         for (int y = rows.Min; y < rows.Max; y++)
                         {
                             Span<TPixel> targetRowSpan = targetPixels.GetRowSpan(y).Slice(startX);
                             PixelOperations<TPixel>.Instance.ToVector4(configuration, targetRowSpan.Slice(0, length), vectorSpan);
 
-                            for (int x = 0; x < width; x++)
+                            if (preserveAlpha)
+                            {
+                                for (int x = 0; x < width; x++)
+                                {
+                                    DenseMatrixUtils.Convolve3(
+                                        in matrix,
+                                        sourcePixels,
+                                        ref vectorSpanRef,
+                                        y,
+                                        x,
+                                        startY,
+                                        maxY,
+                                        startX,
+                                        maxX);
+                                }
+                            }
+                            else
                             {
-                                DenseMatrixUtils.Convolve(in matrix, sourcePixels, vectorSpan, y, x, maxY, maxX, startX);
+                                for (int x = 0; x < width; x++)
+                                {
+                                    DenseMatrixUtils.Convolve4(
+                                        in matrix,
+                                        sourcePixels,
+                                        ref vectorSpanRef,
+                                        y,
+                                        x,
+                                        startY,
+                                        maxY,
+                                        startX,
+                                        maxX);
+                                }
                             }
 
-                            PixelOperations<TPixel>.Instance.FromVector4Destructive(configuration, vectorSpan.Slice(0, length), targetRowSpan);
+                            PixelOperations<TPixel>.Instance.FromVector4Destructive(configuration, vectorSpan, targetRowSpan);
                         }
                     });
         }
diff --git a/src/ImageSharp/Processing/Processors/Convolution/ConvolutionProcessor.cs b/src/ImageSharp/Processing/Processors/Convolution/ConvolutionProcessor.cs
index 7e003cb03..6c3b9a46f 100644
--- a/src/ImageSharp/Processing/Processors/Convolution/ConvolutionProcessor.cs
+++ b/src/ImageSharp/Processing/Processors/Convolution/ConvolutionProcessor.cs
@@ -3,6 +3,7 @@
 
 using System;
 using System.Numerics;
+using System.Runtime.InteropServices;
 using SixLabors.ImageSharp.Memory;
 using SixLabors.ImageSharp.ParallelUtils;
 using SixLabors.ImageSharp.PixelFormats;
@@ -22,17 +23,29 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         /// Initializes a new instance of the <see cref="ConvolutionProcessor{TPixel}"/> class.
         /// </summary>
         /// <param name="kernelXY">The 2d gradient operator.</param>
-        public ConvolutionProcessor(DenseMatrix<float> kernelXY) => this.KernelXY = kernelXY;
+        /// <param name="preserveAlpha">Whether the convolution filter is applied to alpha as well as the color channels.</param>
+        public ConvolutionProcessor(in DenseMatrix<float> kernelXY, bool preserveAlpha)
+        {
+            this.KernelXY = kernelXY;
+            this.PreserveAlpha = preserveAlpha;
+        }
 
         /// <summary>
         /// Gets the 2d gradient operator.
         /// </summary>
         public DenseMatrix<float> KernelXY { get; }
 
+        /// <summary>
+        /// Gets a value indicating whether the convolution filter is applied to alpha as well as the color channels.
+        /// </summary>
+        public bool PreserveAlpha { get; }
+
         /// <inheritdoc/>
         protected override void OnFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration)
         {
             DenseMatrix<float> matrix = this.KernelXY;
+            bool preserveAlpha = this.PreserveAlpha;
+
             var interest = Rectangle.Intersect(sourceRectangle, source.Bounds());
             int startY = interest.Y;
             int endY = interest.Bottom;
@@ -55,18 +68,47 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
                         {
                             Span<Vector4> vectorSpan = vectorBuffer.Span;
                             int length = vectorSpan.Length;
+                            ref Vector4 vectorSpanRef = ref MemoryMarshal.GetReference(vectorSpan);
 
                             for (int y = rows.Min; y < rows.Max; y++)
                             {
                                 Span<TPixel> targetRowSpan = targetPixels.GetRowSpan(y).Slice(startX);
                                 PixelOperations<TPixel>.Instance.ToVector4(configuration, targetRowSpan.Slice(0, length), vectorSpan);
 
-                                for (int x = 0; x < width; x++)
+                                if (preserveAlpha)
+                                {
+                                    for (int x = 0; x < width; x++)
+                                    {
+                                        DenseMatrixUtils.Convolve3(
+                                            in matrix,
+                                            source.PixelBuffer,
+                                            ref vectorSpanRef,
+                                            y,
+                                            x,
+                                            startY,
+                                            maxY,
+                                            startX,
+                                            maxX);
+                                    }
+                                }
+                                else
                                 {
-                                    DenseMatrixUtils.Convolve(in matrix, source.PixelBuffer, vectorSpan, y, x, maxY, maxX, startX);
+                                    for (int x = 0; x < width; x++)
+                                    {
+                                        DenseMatrixUtils.Convolve4(
+                                            in matrix,
+                                            source.PixelBuffer,
+                                            ref vectorSpanRef,
+                                            y,
+                                            x,
+                                            startY,
+                                            maxY,
+                                            startX,
+                                            maxX);
+                                    }
                                 }
 
-                                PixelOperations<TPixel>.Instance.FromVector4Destructive(configuration, vectorSpan.Slice(0, length), targetRowSpan);
+                                PixelOperations<TPixel>.Instance.FromVector4Destructive(configuration, vectorSpan, targetRowSpan);
                             }
                         });
 
diff --git a/src/ImageSharp/Processing/Processors/Convolution/EdgeDetector2DProcessor.cs b/src/ImageSharp/Processing/Processors/Convolution/EdgeDetector2DProcessor.cs
index 892771649..d2e9630dc 100644
--- a/src/ImageSharp/Processing/Processors/Convolution/EdgeDetector2DProcessor.cs
+++ b/src/ImageSharp/Processing/Processors/Convolution/EdgeDetector2DProcessor.cs
@@ -21,7 +21,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         /// <param name="kernelX">The horizontal gradient operator.</param>
         /// <param name="kernelY">The vertical gradient operator.</param>
         /// <param name="grayscale">Whether to convert the image to grayscale before performing edge detection.</param>
-        protected EdgeDetector2DProcessor(DenseMatrix<float> kernelX, DenseMatrix<float> kernelY, bool grayscale)
+        protected EdgeDetector2DProcessor(in DenseMatrix<float> kernelX, in DenseMatrix<float> kernelY, bool grayscale)
         {
             Guard.IsTrue(kernelX.Size.Equals(kernelY.Size), $"{nameof(kernelX)} {nameof(kernelY)}", "Kernel sizes must be the same.");
             this.KernelX = kernelX;
@@ -43,7 +43,8 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         public bool Grayscale { get; set; }
 
         /// <inheritdoc />
-        protected override void OnFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration) => new Convolution2DProcessor<TPixel>(this.KernelX, this.KernelY).Apply(source, sourceRectangle, configuration);
+        protected override void OnFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration)
+            => new Convolution2DProcessor<TPixel>(this.KernelX, this.KernelY, true).Apply(source, sourceRectangle, configuration);
 
         /// <inheritdoc/>
         protected override void BeforeFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration)
diff --git a/src/ImageSharp/Processing/Processors/Convolution/EdgeDetectorCompassProcessor.cs b/src/ImageSharp/Processing/Processors/Convolution/EdgeDetectorCompassProcessor.cs
index 4165cf024..73f92fae3 100644
--- a/src/ImageSharp/Processing/Processors/Convolution/EdgeDetectorCompassProcessor.cs
+++ b/src/ImageSharp/Processing/Processors/Convolution/EdgeDetectorCompassProcessor.cs
@@ -5,14 +5,12 @@ using System;
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
-using System.Threading.Tasks;
 
 using SixLabors.ImageSharp.Memory;
 using SixLabors.ImageSharp.ParallelUtils;
 using SixLabors.ImageSharp.PixelFormats;
 using SixLabors.ImageSharp.Primitives;
 using SixLabors.ImageSharp.Processing.Processors.Filters;
-using SixLabors.Memory;
 using SixLabors.Primitives;
 
 namespace SixLabors.ImageSharp.Processing.Processors.Convolution
@@ -28,10 +26,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         /// Initializes a new instance of the <see cref="EdgeDetectorCompassProcessor{TPixel}"/> class.
         /// </summary>
         /// <param name="grayscale">Whether to convert the image to grayscale before performing edge detection.</param>
-        protected EdgeDetectorCompassProcessor(bool grayscale)
-        {
-            this.Grayscale = grayscale;
-        }
+        protected EdgeDetectorCompassProcessor(bool grayscale) => this.Grayscale = grayscale;
 
         /// <summary>
         /// Gets the North gradient operator
@@ -104,7 +99,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
             // we need a clean copy for each pass to start from
             using (ImageFrame<TPixel> cleanCopy = source.Clone())
             {
-                new ConvolutionProcessor<TPixel>(kernels[0]).Apply(source, sourceRectangle, configuration);
+                new ConvolutionProcessor<TPixel>(kernels[0], true).Apply(source, sourceRectangle, configuration);
 
                 if (kernels.Length == 1)
                 {
@@ -133,7 +128,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
                 {
                     using (ImageFrame<TPixel> pass = cleanCopy.Clone())
                     {
-                        new ConvolutionProcessor<TPixel>(kernels[i]).Apply(pass, sourceRectangle, configuration);
+                        new ConvolutionProcessor<TPixel>(kernels[i], true).Apply(pass, sourceRectangle, configuration);
 
                         Buffer2D<TPixel> passPixels = pass.PixelBuffer;
                         Buffer2D<TPixel> targetPixels = source.PixelBuffer;
@@ -147,10 +142,8 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
                                     {
                                         int offsetY = y - shiftY;
 
-                                        ref TPixel passPixelsBase =
-                                            ref MemoryMarshal.GetReference(passPixels.GetRowSpan(offsetY));
-                                        ref TPixel targetPixelsBase =
-                                            ref MemoryMarshal.GetReference(targetPixels.GetRowSpan(offsetY));
+                                        ref TPixel passPixelsBase = ref MemoryMarshal.GetReference(passPixels.GetRowSpan(offsetY));
+                                        ref TPixel targetPixelsBase = ref MemoryMarshal.GetReference(targetPixels.GetRowSpan(offsetY));
 
                                         for (int x = minX; x < maxX; x++)
                                         {
@@ -158,8 +151,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
 
                                             // Grab the max components of the two pixels
                                             ref TPixel currentPassPixel = ref Unsafe.Add(ref passPixelsBase, offsetX);
-                                            ref TPixel currentTargetPixel =
-                                                ref Unsafe.Add(ref targetPixelsBase, offsetX);
+                                            ref TPixel currentTargetPixel = ref Unsafe.Add(ref targetPixelsBase, offsetX);
 
                                             var pixelValue = Vector4.Max(
                                                 currentPassPixel.ToVector4(),
diff --git a/src/ImageSharp/Processing/Processors/Convolution/EdgeDetectorProcessor.cs b/src/ImageSharp/Processing/Processors/Convolution/EdgeDetectorProcessor.cs
index 9173bc229..edc7ec4cc 100644
--- a/src/ImageSharp/Processing/Processors/Convolution/EdgeDetectorProcessor.cs
+++ b/src/ImageSharp/Processing/Processors/Convolution/EdgeDetectorProcessor.cs
@@ -20,7 +20,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
         /// </summary>
         /// <param name="kernelXY">The 2d gradient operator.</param>
         /// <param name="grayscale">Whether to convert the image to grayscale before performing edge detection.</param>
-        protected EdgeDetectorProcessor(DenseMatrix<float> kernelXY, bool grayscale)
+        protected EdgeDetectorProcessor(in DenseMatrix<float> kernelXY, bool grayscale)
         {
             this.KernelXY = kernelXY;
             this.Grayscale = grayscale;
@@ -45,8 +45,6 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
 
         /// <inheritdoc/>
         protected override void OnFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration)
-        {
-            new ConvolutionProcessor<TPixel>(this.KernelXY).Apply(source, sourceRectangle, configuration);
-        }
+            => new ConvolutionProcessor<TPixel>(this.KernelXY, true).Apply(source, sourceRectangle, configuration);
     }
 }
\ No newline at end of file
diff --git a/src/ImageSharp/Processing/Processors/Convolution/GaussianBlurProcessor.cs b/src/ImageSharp/Processing/Processors/Convolution/GaussianBlurProcessor.cs
index b3bc15d39..0fc822d57 100644
--- a/src/ImageSharp/Processing/Processors/Convolution/GaussianBlurProcessor.cs
+++ b/src/ImageSharp/Processing/Processors/Convolution/GaussianBlurProcessor.cs
@@ -77,7 +77,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
 
         /// <inheritdoc/>
         protected override void OnFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration)
-            => new Convolution2PassProcessor<TPixel>(this.KernelX, this.KernelY).Apply(source, sourceRectangle, configuration);
+            => new Convolution2PassProcessor<TPixel>(this.KernelX, this.KernelY, false).Apply(source, sourceRectangle, configuration);
 
         /// <summary>
         /// Create a 1 dimensional Gaussian kernel using the Gaussian G(x) function
diff --git a/src/ImageSharp/Processing/Processors/Convolution/GaussianSharpenProcessor.cs b/src/ImageSharp/Processing/Processors/Convolution/GaussianSharpenProcessor.cs
index 786bf7757..001471720 100644
--- a/src/ImageSharp/Processing/Processors/Convolution/GaussianSharpenProcessor.cs
+++ b/src/ImageSharp/Processing/Processors/Convolution/GaussianSharpenProcessor.cs
@@ -79,7 +79,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
 
         /// <inheritdoc/>
         protected override void OnFrameApply(ImageFrame<TPixel> source, Rectangle sourceRectangle, Configuration configuration)
-            => new Convolution2PassProcessor<TPixel>(this.KernelX, this.KernelY).Apply(source, sourceRectangle, configuration);
+            => new Convolution2PassProcessor<TPixel>(this.KernelX, this.KernelY, false).Apply(source, sourceRectangle, configuration);
 
         /// <summary>
         /// Create a 1 dimensional Gaussian kernel using the Gaussian G(x) function
diff --git a/src/ImageSharp/Processing/Processors/Dithering/ErrorDiffuserBase.cs b/src/ImageSharp/Processing/Processors/Dithering/ErrorDiffuserBase.cs
index 642da2f00..abf5dce18 100644
--- a/src/ImageSharp/Processing/Processors/Dithering/ErrorDiffuserBase.cs
+++ b/src/ImageSharp/Processing/Processors/Dithering/ErrorDiffuserBase.cs
@@ -45,7 +45,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Dithering
         /// </summary>
         /// <param name="matrix">The dithering matrix.</param>
         /// <param name="divisor">The divisor.</param>
-        internal ErrorDiffuserBase(DenseMatrix<float> matrix, byte divisor)
+        internal ErrorDiffuserBase(in DenseMatrix<float> matrix, byte divisor)
         {
             Guard.MustBeGreaterThan(divisor, 0, nameof(divisor));
 
diff --git a/tests/ImageSharp.Benchmarks/Samplers/GaussianBlur.cs b/tests/ImageSharp.Benchmarks/Samplers/GaussianBlur.cs
new file mode 100644
index 000000000..47bd42a3c
--- /dev/null
+++ b/tests/ImageSharp.Benchmarks/Samplers/GaussianBlur.cs
@@ -0,0 +1,19 @@
+using BenchmarkDotNet.Attributes;
+using SixLabors.ImageSharp.PixelFormats;
+using SixLabors.ImageSharp.Processing;
+
+namespace SixLabors.ImageSharp.Benchmarks.Samplers
+{
+    [Config(typeof(Config.ShortClr))]
+    public class GaussianBlur
+    {
+        [Benchmark]
+        public void Blur()
+        {
+            using (var image = new Image<Rgba32>(Configuration.Default, 400, 400, Rgba32.White))
+            {
+                image.Mutate(c => c.GaussianBlur());
+            }
+        }
+    }
+}
diff --git a/tests/ImageSharp.Tests/MetaData/Profiles/ICC/DataWriter/IccDataWriter.MatrixTests.cs b/tests/ImageSharp.Tests/MetaData/Profiles/ICC/DataWriter/IccDataWriter.MatrixTests.cs
index b43c17f79..088222222 100644
--- a/tests/ImageSharp.Tests/MetaData/Profiles/ICC/DataWriter/IccDataWriter.MatrixTests.cs
+++ b/tests/ImageSharp.Tests/MetaData/Profiles/ICC/DataWriter/IccDataWriter.MatrixTests.cs
@@ -38,7 +38,7 @@ namespace SixLabors.ImageSharp.Tests.Icc
 
         [Theory]
         [MemberData(nameof(IccTestDataMatrix.Matrix2D_DenseMatrixTestData), MemberType = typeof(IccTestDataMatrix))]
-        internal void WriteMatrix2D_DenseMatrix(byte[] expected, int xCount, int yCount, bool isSingle, DenseMatrix<float> data)
+        internal void WriteMatrix2D_DenseMatrix(byte[] expected, int xCount, int yCount, bool isSingle, in DenseMatrix<float> data)
         {
             IccDataWriter writer = CreateWriter();
 
diff --git a/tests/ImageSharp.Tests/Processing/Processors/Convolution/BoxBlurTest.cs b/tests/ImageSharp.Tests/Processing/Processors/Convolution/BoxBlurTest.cs
index 0c40debad..1f0d12cf7 100644
--- a/tests/ImageSharp.Tests/Processing/Processors/Convolution/BoxBlurTest.cs
+++ b/tests/ImageSharp.Tests/Processing/Processors/Convolution/BoxBlurTest.cs
@@ -37,7 +37,7 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Convolution
             where TPixel : struct, IPixel<TPixel>
         {
             using (Image<TPixel> source = provider.GetImage())
-            using (var image = source.Clone())
+            using (Image<TPixel> image = source.Clone())
             {
                 var bounds = new Rectangle(10, 10, image.Width / 2, image.Height / 2);
 
diff --git a/tests/ImageSharp.Tests/Processing/Processors/Convolution/DetectEdgesTest.cs b/tests/ImageSharp.Tests/Processing/Processors/Convolution/DetectEdgesTest.cs
index edb24d6f1..b6a7741b3 100644
--- a/tests/ImageSharp.Tests/Processing/Processors/Convolution/DetectEdgesTest.cs
+++ b/tests/ImageSharp.Tests/Processing/Processors/Convolution/DetectEdgesTest.cs
@@ -12,6 +12,8 @@ namespace SixLabors.ImageSharp.Tests.Processing.Processors.Convolution
 {
     public class DetectEdgesTest : FileTestBase
     {
+        // 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);
 
         public static readonly string[] CommonTestImages = { TestImages.Png.Bike };