tsai
/
ImageSharp
mirror of https://github.com/SixLabors/ImageSharp
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
Browse Source

Specialize bokeh blur operations for 1D kernels

js/color-alpha-handling
Sergio Pedri 5 years ago
parent
b3f4befe5e
commit
ca1a67a36c
2 changed files with 40 additions and 40 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 34
      src/ImageSharp/Processing/Processors/Convolution/BokehBlurProcessor.cs
  2. 46
      src/ImageSharp/Processing/Processors/Convolution/BokehBlurProcessor{TPixel}.cs

34
src/ImageSharp/Processing/Processors/Convolution/BokehBlurProcessor.cs
View File

@ -127,39 +127,37 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
{
int boundsX = this.bounds.X;
int boundsWidth = this.bounds.Width;
Span<Vector4> targetBuffer = this.targetValues.GetRowSpan(y);
int kernelSize = this.kernel.Length;
var state = new ConvolutionState<Complex64>(this.kernel, this.kernel.Length, 1, this.map);
ref int sampleRowBase = ref state.GetSampleRow(y - this.bounds.Y);
Span<int> rowOffsets = this.map.GetRowOffsetSpan();
Span<int> columnOffsets = this.map.GetColumnOffsetSpan();
ref int sampleRowBase = ref Unsafe.Add(ref MemoryMarshal.GetReference(rowOffsets), (y - this.bounds.Y) * kernelSize);
ref int sampleColumnBase = ref MemoryMarshal.GetReference(columnOffsets);
// The target buffer is zeroed initially and then it accumulates the results
// of each partial convolution, so we don't have to clear it here as well.
ref Vector4 targetBase = ref MemoryMarshal.GetReference(targetBuffer);
// of each partial convolution, so we don't have to clear it here as well
Span<Vector4> targetBuffer = this.targetValues.GetRowSpan(y);
ReadOnlyKernel<Complex64> kernel = state.Kernel;
ref Vector4 targetBase = ref MemoryMarshal.GetReference(targetBuffer);
ref Complex64 kernelBase = ref this.kernel[0];
for (int kY = 0; kY < kernel.Rows; kY++)
for (int kY = 0; kY < kernelSize; kY++)
{
// Get the precalculated source sample row for this kernel row and copy to our buffer.
// Get the precalculated source sample row for this kernel row and copy to our buffer
int sampleY = Unsafe.Add(ref sampleRowBase, kY);
Span<ComplexVector4> sourceRow = this.sourceValues.GetRowSpan(sampleY).Slice(boundsX, boundsWidth);
ref ComplexVector4 sourceBase = ref MemoryMarshal.GetReference(sourceRow);
Complex64 factor = Unsafe.Add(ref kernelBase, kY);
for (int x = 0; x < boundsWidth; x++)
{
ref int sampleColumnBase = ref state.GetSampleColumn(x);
int sampleX = Unsafe.Add(ref sampleColumnBase, x) - boundsX;
ref Vector4 target = ref Unsafe.Add(ref targetBase, x);
ComplexVector4 pixel4 = default;
for (int kX = 0; kX < kernel.Columns; kX++)
{
int sampleX = Unsafe.Add(ref sampleColumnBase, kX) - boundsX;
ComplexVector4 sample = Unsafe.Add(ref sourceBase, sampleX);
pixel4.Sum(kernel[kY, kX] * sample);
}
ComplexVector4 sample = Unsafe.Add(ref sourceBase, sampleX);
ComplexVector4 partial = factor * sample;
target += pixel4.WeightedSum(this.z, this.w);
target += partial.WeightedSum(this.z, this.w);
}
}
}

46
src/ImageSharp/Processing/Processors/Convolution/BokehBlurProcessor{TPixel}.cs
View File

@ -207,39 +207,41 @@ namespace SixLabors.ImageSharp.Processing.Processors.Convolution
{
int boundsX = this.bounds.X;
int boundsWidth = this.bounds.Width;
int kernelSize = this.kernel.Length;
var state = new ConvolutionState<Complex64>(this.kernel, 1, this.kernel.Length, this.map);
ref int sampleRowBase = ref state.GetSampleRow(y - this.bounds.Y);
Span<int> rowOffsets = this.map.GetRowOffsetSpan();
Span<int> columnOffsets = this.map.GetColumnOffsetSpan();
int sampleY = Unsafe.Add(ref MemoryMarshal.GetReference(rowOffsets), y - this.bounds.Y);
ref int sampleColumnBase = ref MemoryMarshal.GetReference(columnOffsets);
// Clear the target buffer for each row run
Span<ComplexVector4> targetBuffer = this.targetValues.GetRowSpan(y);
// Clear the target buffer
targetBuffer.Clear();
ref ComplexVector4 targetBase = ref MemoryMarshal.GetReference(targetBuffer);
ReadOnlyKernel<Complex64> kernel = state.Kernel;
// Execute the bulk pixel format conversion for the current row
Span<TPixel> sourceRow = this.sourcePixels.GetRowSpan(sampleY).Slice(boundsX, boundsWidth);
PixelOperations<TPixel>.Instance.ToVector4(this.configuration, sourceRow, span);
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, span);
ref Vector4 sourceBase = ref MemoryMarshal.GetReference(span);
ref Complex64 kernelBase = ref this.kernel[0];
ref Vector4 sourceBase = ref MemoryMarshal.GetReference(span);
for (int x = 0; x < span.Length; x++)
{
ref ComplexVector4 target = ref Unsafe.Add(ref targetBase, x);
for (int x = 0; x < span.Length; x++)
for (int kX = 0; kX < kernelSize; kX++)
{
ref int sampleColumnBase = ref state.GetSampleColumn(x);
ref ComplexVector4 target = ref Unsafe.Add(ref targetBase, x);
for (int kX = 0; kX < kernel.Columns; kX++)
{
int sampleX = Unsafe.Add(ref sampleColumnBase, kX) - boundsX;
Vector4 sample = Unsafe.Add(ref sourceBase, sampleX);
target.Sum(kernel[kY, kX] * sample);
}
int sampleX = Unsafe.Add(ref sampleColumnBase, kX) - boundsX;
Vector4 sample = Unsafe.Add(ref sourceBase, sampleX);
Complex64 factor = Unsafe.Add(ref kernelBase, kX);
target.Sum(factor * sample);
}
// Shift the base column sampling reference by one row at the end of each outer
// iteration so that the inner tight loop indexing can skip the multiplication
sampleColumnBase = ref Unsafe.Add(ref sampleColumnBase, kernelSize);
}
}
}

Write Preview
Loading…
Cancel
Save