Browse Source

Port vertical convolution processor, remove X loop

js/color-alpha-handling
Sergio Pedri 5 years ago
parent
commit
d41cf15023
  1. 147
      src/ImageSharp/Processing/Processors/Convolution/Convolution2PassProcessor{TPixel}.cs

147
src/ImageSharp/Processing/Processors/Convolution/Convolution2PassProcessor{TPixel}.cs

@ -92,7 +92,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
mapY.BuildSamplingOffsetMap(this.KernelY, interest);
// Vertical convolution
var verticalOperation = new ConvolutionRowOperation<TPixel>(
var verticalOperation = new VerticalConvolutionRowOperation(
interest,
source.PixelBuffer,
firstPassPixels,
@ -101,7 +101,7 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
this.Configuration,
this.PreserveAlpha);
ParallelRowIterator.IterateRows<ConvolutionRowOperation<TPixel>, Vector4>(
ParallelRowIterator.IterateRows<VerticalConvolutionRowOperation, Vector4>(
this.Configuration,
operationBounds,
in verticalOperation);
@ -249,5 +249,148 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
PixelOperations<TPixel>.Instance.FromVector4Destructive(this.configuration, targetBuffer, targetRow);
}
}
/// <summary>
/// A <see langword="struct"/> implementing the logic for the vertical 1D convolution.
/// </summary>
internal readonly struct VerticalConvolutionRowOperation : IRowOperation<Vector4>
{
private readonly Rectangle bounds;
private readonly Buffer2D<TPixel> targetPixels;
private readonly Buffer2D<TPixel> sourcePixels;
private readonly KernelSamplingMap map;
private readonly DenseMatrix<float> kernelMatrix;
private readonly Configuration configuration;
private readonly bool preserveAlpha;
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public VerticalConvolutionRowOperation(
Rectangle bounds,
Buffer2D<TPixel> targetPixels,
Buffer2D<TPixel> sourcePixels,
KernelSamplingMap map,
DenseMatrix<float> kernelMatrix,
Configuration configuration,
bool preserveAlpha)
{
this.bounds = bounds;
this.targetPixels = targetPixels;
this.sourcePixels = sourcePixels;
this.map = map;
this.kernelMatrix = kernelMatrix;
this.configuration = configuration;
this.preserveAlpha = preserveAlpha;
}
/// <inheritdoc/>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public void Invoke(int y, Span<Vector4> span)
{
if (this.preserveAlpha)
{
this.Convolve3(y, span);
}
else
{
this.Convolve4(y, span);
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private void Convolve3(int y, Span<Vector4> span)
{
// Span is 2x bounds.
int boundsX = this.bounds.X;
int boundsWidth = this.bounds.Width;
Span<Vector4> sourceBuffer = span.Slice(0, this.bounds.Width);
Span<Vector4> targetBuffer = span.Slice(this.bounds.Width);
var state = new ConvolutionState(in this.kernelMatrix, this.map);
ref int sampleRowBase = ref state.GetSampleRow(y - this.bounds.Y);
// Clear the target buffer for each row run.
targetBuffer.Clear();
ref Vector4 targetBase = ref MemoryMarshal.GetReference(targetBuffer);
ReadOnlyKernel kernel = state.Kernel;
Span<TPixel> sourceRow;
for (int kY = 0; kY < kernel.Rows; kY++)
{
// Get the precalculated source sample row for this kernel row and copy to our buffer.
int sampleY = Unsafe.Add(ref sampleRowBase, kY);
sourceRow = this.sourcePixels.GetRowSpan(sampleY).Slice(boundsX, boundsWidth);
PixelOperations<TPixel>.Instance.ToVector4(this.configuration, sourceRow, sourceBuffer);
ref Vector4 sourceBase = ref MemoryMarshal.GetReference(sourceBuffer);
for (int x = 0; x < sourceBuffer.Length; x++)
{
ref int sampleColumnBase = ref state.GetSampleColumn(x);
ref Vector4 target = ref Unsafe.Add(ref targetBase, x);
int sampleX = Unsafe.Add(ref sampleColumnBase, 0) - boundsX;
Vector4 sample = Unsafe.Add(ref sourceBase, sampleX);
target += kernel[kY, 0] * sample;
}
}
// Now we need to copy the original alpha values from the source row.
sourceRow = this.sourcePixels.GetRowSpan(y).Slice(boundsX, boundsWidth);
PixelOperations<TPixel>.Instance.ToVector4(this.configuration, sourceRow, sourceBuffer);
for (int x = 0; x < sourceRow.Length; x++)
{
ref Vector4 target = ref Unsafe.Add(ref targetBase, x);
target.W = Unsafe.Add(ref MemoryMarshal.GetReference(sourceBuffer), x).W;
}
Span<TPixel> targetRow = this.targetPixels.GetRowSpan(y).Slice(boundsX, boundsWidth);
PixelOperations<TPixel>.Instance.FromVector4Destructive(this.configuration, targetBuffer, targetRow);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private void Convolve4(int y, Span<Vector4> span)
{
// Span is 2x bounds.
int boundsX = this.bounds.X;
int boundsWidth = this.bounds.Width;
Span<Vector4> sourceBuffer = span.Slice(0, this.bounds.Width);
Span<Vector4> targetBuffer = span.Slice(this.bounds.Width);
var state = new ConvolutionState(in this.kernelMatrix, this.map);
ref int sampleRowBase = ref state.GetSampleRow(y - this.bounds.Y);
// Clear the target buffer for each row run.
targetBuffer.Clear();
ref Vector4 targetBase = ref MemoryMarshal.GetReference(targetBuffer);
ReadOnlyKernel kernel = state.Kernel;
for (int kY = 0; kY < kernel.Rows; kY++)
{
// Get the precalculated source sample row for this kernel row and copy to our buffer.
int sampleY = Unsafe.Add(ref sampleRowBase, kY);
Span<TPixel> sourceRow = this.sourcePixels.GetRowSpan(sampleY).Slice(boundsX, boundsWidth);
PixelOperations<TPixel>.Instance.ToVector4(this.configuration, sourceRow, sourceBuffer);
Numerics.Premultiply(sourceBuffer);
ref Vector4 sourceBase = ref MemoryMarshal.GetReference(sourceBuffer);
for (int x = 0; x < sourceBuffer.Length; x++)
{
ref int sampleColumnBase = ref state.GetSampleColumn(x);
ref Vector4 target = ref Unsafe.Add(ref targetBase, x);
int sampleX = Unsafe.Add(ref sampleColumnBase, 0) - boundsX;
Vector4 sample = Unsafe.Add(ref sourceBase, sampleX);
target += kernel[kY, 0] * sample;
}
}
Numerics.UnPremultiply(targetBuffer);
Span<TPixel> targetRow = this.targetPixels.GetRowSpan(y).Slice(boundsX, boundsWidth);
PixelOperations<TPixel>.Instance.FromVector4Destructive(this.configuration, targetBuffer, targetRow);
}
}
}
}

Loading…
Cancel
Save