Replace parallel row iteration with sequential loops in Crop/Convolution Processors

src/ImageSharp/Processing/Processors/Convolution/BokehBlurProcessor{TPixel}.cs

@@ -5,6 +5,7 @@
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 using SixLabors.ImageSharp.Advanced;
+using SixLabors.ImageSharp.ColorProfiles.Companding;
 using SixLabors.ImageSharp.Memory;
 using SixLabors.ImageSharp.PixelFormats;
 using SixLabors.ImageSharp.Processing.Processors.Convolution.Parameters;
@@ -112,7 +113,7 @@ protected override void OnFrameApply(ImageFrame<TPixel> source)
         }
         else
         {
-            ApplyInverseGammaExposureRowOperation operation = new(sourceRectangle, source.PixelBuffer, processingBuffer, this.Configuration, 1 / this.gamma);
+            ApplyInverseGammaExposureRowOperation operation = new(sourceRectangle, source.PixelBuffer, processingBuffer, this.Configuration, this.gamma);
             ParallelRowIterator.IterateRows(
                 this.Configuration,
                 sourceRectangle,
@@ -305,15 +306,9 @@ public void Invoke(int y, Span<Vector4> span)
         {
             Span<TPixel> targetRowSpan = this.targetPixels.DangerousGetRowSpan(y)[this.bounds.X..];
             PixelOperations<TPixel>.Instance.ToVector4(this.configuration, targetRowSpan[..span.Length], span, PixelConversionModifiers.Premultiply);
-            ref Vector4 baseRef = ref MemoryMarshal.GetReference(span);
 
-            for (int x = 0; x < this.bounds.Width; x++)
-            {
-                ref Vector4 v = ref Unsafe.Add(ref baseRef, (uint)x);
-                v.X = MathF.Pow(v.X, this.gamma);
-                v.Y = MathF.Pow(v.Y, this.gamma);
-                v.Z = MathF.Pow(v.Z, this.gamma);
-            }
+            // Input is premultiplied [0,1] so the LUT is safe here.
+            GammaCompanding.Expand(span[..this.bounds.Width], this.gamma);
 
             PixelOperations<TPixel>.Instance.FromVector4Destructive(this.configuration, span, targetRowSpan);
         }
@@ -367,44 +362,34 @@ public void Invoke(int y, Span<Vector4> span)
         private readonly Buffer2D<TPixel> targetPixels;
         private readonly Buffer2D<Vector4> sourceValues;
         private readonly Configuration configuration;
-        private readonly float inverseGamma;
+        private readonly float gamma;
 
         [MethodImpl(InliningOptions.ShortMethod)]
         public ApplyInverseGammaExposureRowOperation(
             Rectangle bounds,
             Buffer2D<TPixel> targetPixels,
             Buffer2D<Vector4> sourceValues,
             Configuration configuration,
-            float inverseGamma)
+            float gamma)
         {
             this.bounds = bounds;
             this.targetPixels = targetPixels;
             this.sourceValues = sourceValues;
             this.configuration = configuration;
-            this.inverseGamma = inverseGamma;
+            this.gamma = gamma;
         }
 
         /// <inheritdoc/>
         [MethodImpl(InliningOptions.ShortMethod)]
         public void Invoke(int y)
         {
-            Vector4 low = Vector4.Zero;
-            Vector4 high = new(float.PositiveInfinity, float.PositiveInfinity, float.PositiveInfinity, float.PositiveInfinity);
-
             Span<TPixel> targetPixelSpan = this.targetPixels.DangerousGetRowSpan(y)[this.bounds.X..];
-            Span<Vector4> sourceRowSpan = this.sourceValues.DangerousGetRowSpan(y)[this.bounds.X..];
-            ref Vector4 sourceRef = ref MemoryMarshal.GetReference(sourceRowSpan);
+            Span<Vector4> sourceRowSpan = this.sourceValues.DangerousGetRowSpan(y).Slice(this.bounds.X, this.bounds.Width);
 
-            for (int x = 0; x < this.bounds.Width; x++)
-            {
-                ref Vector4 v = ref Unsafe.Add(ref sourceRef, (uint)x);
-                Vector4 clamp = Numerics.Clamp(v, low, high);
-                v.X = MathF.Pow(clamp.X, this.inverseGamma);
-                v.Y = MathF.Pow(clamp.Y, this.inverseGamma);
-                v.Z = MathF.Pow(clamp.Z, this.inverseGamma);
-            }
+            Numerics.Clamp(MemoryMarshal.Cast<Vector4, float>(sourceRowSpan), 0, 1F);
+            GammaCompanding.Compress(sourceRowSpan, this.gamma);
 
-            PixelOperations<TPixel>.Instance.FromVector4Destructive(this.configuration, sourceRowSpan[..this.bounds.Width], targetPixelSpan, PixelConversionModifiers.Premultiply);
+            PixelOperations<TPixel>.Instance.FromVector4Destructive(this.configuration, sourceRowSpan, targetPixelSpan, PixelConversionModifiers.Premultiply);
         }
     }
 
@@ -433,17 +418,16 @@ public ApplyInverseGamma3ExposureRowOperation(
 
         /// <inheritdoc/>
         [MethodImpl(InliningOptions.ShortMethod)]
-        public unsafe void Invoke(int y)
+        public void Invoke(int y)
         {
             Span<Vector4> sourceRowSpan = this.sourceValues.DangerousGetRowSpan(y).Slice(this.bounds.X, this.bounds.Width);
-            ref Vector4 sourceRef = ref MemoryMarshal.GetReference(sourceRowSpan);
 
-            Numerics.Clamp(MemoryMarshal.Cast<Vector4, float>(sourceRowSpan), 0, float.PositiveInfinity);
+            Numerics.Clamp(MemoryMarshal.Cast<Vector4, float>(sourceRowSpan), 0, 1F);
             Numerics.CubeRootOnXYZ(sourceRowSpan);
 
             Span<TPixel> targetPixelSpan = this.targetPixels.DangerousGetRowSpan(y)[this.bounds.X..];
 
-            PixelOperations<TPixel>.Instance.FromVector4Destructive(this.configuration, sourceRowSpan[..this.bounds.Width], targetPixelSpan, PixelConversionModifiers.Premultiply);
+            PixelOperations<TPixel>.Instance.FromVector4Destructive(this.configuration, sourceRowSpan, targetPixelSpan, PixelConversionModifiers.Premultiply);
         }
     }
 }

src/ImageSharp/Processing/Processors/Convolution/Convolution2DProcessor{TPixel}.cs

@@ -1,8 +1,8 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
 using System.Numerics;
-using SixLabors.ImageSharp.Advanced;
 using SixLabors.ImageSharp.Memory;
 using SixLabors.ImageSharp.PixelFormats;
 
@@ -79,10 +79,16 @@ protected override void OnFrameApply(ImageFrame<TPixel> source)
                 this.Configuration,
                 this.PreserveAlpha);
 
-            ParallelRowIterator.IterateRows<Convolution2DRowOperation<TPixel>, Vector4>(
-                this.Configuration,
-                interest,
-                in operation);
+            // Convolution is memory-bandwidth-bound with low arithmetic intensity.
+            // Parallelization degrades performance due to cache line contention from
+            // overlapping source row reads. See #3111.
+            using IMemoryOwner<Vector4> buffer = allocator.Allocate<Vector4>(operation.GetRequiredBufferLength(interest));
+            Span<Vector4> span = buffer.Memory.Span;
+
+            for (int y = interest.Top; y < interest.Bottom; y++)
+            {
+                operation.Invoke(y, span);
+            }
         }
 
         Buffer2D<TPixel>.SwapOrCopyContent(source.PixelBuffer, targetPixels);

src/ImageSharp/Processing/Processors/Convolution/Convolution2PassProcessor{TPixel}.cs

@@ -1,6 +1,7 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
@@ -106,6 +107,12 @@ protected override void OnFrameApply(ImageFrame<TPixel> source)
 
         mapXY.BuildSamplingOffsetMap(this.KernelX.Length, this.KernelX.Length, interest, this.BorderWrapModeX, this.BorderWrapModeY);
 
+        MemoryAllocator allocator = this.Configuration.MemoryAllocator;
+
+        // Convolution is memory-bandwidth-bound with low arithmetic intensity.
+        // Parallelization degrades performance due to cache line contention from
+        // overlapping source row reads. See #3111.
+
         // Horizontal convolution
         HorizontalConvolutionRowOperation horizontalOperation = new(
             interest,
@@ -116,10 +123,15 @@ protected override void OnFrameApply(ImageFrame<TPixel> source)
             this.Configuration,
             this.PreserveAlpha);
 
-        ParallelRowIterator.IterateRows<HorizontalConvolutionRowOperation, Vector4>(
-            this.Configuration,
-            interest,
-            in horizontalOperation);
+        using (IMemoryOwner<Vector4> hBuffer = allocator.Allocate<Vector4>(horizontalOperation.GetRequiredBufferLength(interest)))
+        {
+            Span<Vector4> hSpan = hBuffer.Memory.Span;
+
+            for (int y = interest.Top; y < interest.Bottom; y++)
+            {
+                horizontalOperation.Invoke(y, hSpan);
+            }
+        }
 
         // Vertical convolution
         VerticalConvolutionRowOperation verticalOperation = new(
@@ -131,10 +143,15 @@ protected override void OnFrameApply(ImageFrame<TPixel> source)
             this.Configuration,
             this.PreserveAlpha);
 
-        ParallelRowIterator.IterateRows<VerticalConvolutionRowOperation, Vector4>(
-            this.Configuration,
-            interest,
-            in verticalOperation);
+        using (IMemoryOwner<Vector4> vBuffer = allocator.Allocate<Vector4>(verticalOperation.GetRequiredBufferLength(interest)))
+        {
+            Span<Vector4> vSpan = vBuffer.Memory.Span;
+
+            for (int y = interest.Top; y < interest.Bottom; y++)
+            {
+                verticalOperation.Invoke(y, vSpan);
+            }
+        }
     }
 
     /// <summary>

src/ImageSharp/Processing/Processors/Convolution/ConvolutionProcessor{TPixel}.cs

@@ -1,6 +1,7 @@
 // Copyright (c) Six Labors.
 // Licensed under the Six Labors Split License.
 
+using System.Buffers;
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
@@ -96,10 +97,17 @@ protected override void OnFrameApply(ImageFrame<TPixel> source)
             map.BuildSamplingOffsetMap(this.KernelXY.Rows, this.KernelXY.Columns, interest, this.BorderWrapModeX, this.BorderWrapModeY);
 
             RowOperation operation = new(interest, targetPixels, source.PixelBuffer, map, this.KernelXY, this.Configuration, this.PreserveAlpha);
-            ParallelRowIterator.IterateRows<RowOperation, Vector4>(
-               this.Configuration,
-               interest,
-               in operation);
+
+            // Convolution is memory-bandwidth-bound with low arithmetic intensity.
+            // Parallelization degrades performance due to cache line contention from
+            // overlapping source row reads. See #3111.
+            using IMemoryOwner<Vector4> buffer = allocator.Allocate<Vector4>(operation.GetRequiredBufferLength(interest));
+            Span<Vector4> span = buffer.Memory.Span;
+
+            for (int y = interest.Top; y < interest.Bottom; y++)
+            {
+                operation.Invoke(y, span);
+            }
         }
 
         Buffer2D<TPixel>.SwapOrCopyContent(source.PixelBuffer, targetPixels);

src/ImageSharp/Processing/Processors/Convolution/EdgeDetectorCompassProcessor{TPixel}.cs

@@ -83,11 +83,15 @@ protected override void OnFrameApply(ImageFrame<TPixel> source)
                 processor.Apply(pass);
             }
 
+            // Convolution is memory-bandwidth-bound with low arithmetic intensity.
+            // Parallelization degrades performance due to cache line contention from
+            // overlapping source row reads. See #3111.
             RowOperation operation = new(source.PixelBuffer, pass.PixelBuffer, interest);
-            ParallelRowIterator.IterateRows(
-                this.Configuration,
-                interest,
-                in operation);
+
+            for (int y = interest.Top; y < interest.Bottom; y++)
+            {
+                operation.Invoke(y);
+            }
         }
     }
 

src/ImageSharp/Processing/Processors/Transforms/CropProcessor{TPixel}.cs

@@ -59,16 +59,14 @@ protected override void OnFrameApply(ImageFrame<TPixel> source, ImageFrame<TPixe
 
         Rectangle bounds = this.cropRectangle;
 
-        // Copying is cheap, we should process more pixels per task:
-        ParallelExecutionSettings parallelSettings =
-            ParallelExecutionSettings.FromConfiguration(this.Configuration).MultiplyMinimumPixelsPerTask(4);
-
+        // Copying is too cheap to benefit from parallelization;
+        // the overhead exceeds the work per task. See #3111.
         RowOperation operation = new(bounds, source.PixelBuffer, destination.PixelBuffer);
 
-        ParallelRowIterator.IterateRows(
-            bounds,
-            in parallelSettings,
-            in operation);
+        for (int y = bounds.Top; y < bounds.Bottom; y++)
+        {
+            operation.Invoke(y);
+        }
     }
 
     /// <summary>