Browse Source

Removed obsolete code

pull/2120/head
Dmitry Pentin 4 years ago
parent
commit
a5305ba5ca
  1. 35
      src/ImageSharp/Formats/Jpeg/Components/Encoder/HuffIndex.cs
  2. 318
      src/ImageSharp/Formats/Jpeg/Components/Encoder/HuffmanScanEncoder.cs
  3. 127
      src/ImageSharp/Formats/Jpeg/Components/Encoder/LuminanceForwardConverter{TPixel}.cs
  4. 165
      src/ImageSharp/Formats/Jpeg/Components/Encoder/RgbForwardConverter{TPixel}.cs
  5. 237
      src/ImageSharp/Formats/Jpeg/Components/Encoder/RgbToYCbCrConverterLut.cs
  6. 259
      src/ImageSharp/Formats/Jpeg/Components/Encoder/RgbToYCbCrConverterVectorized.cs
  7. 121
      src/ImageSharp/Formats/Jpeg/Components/Encoder/YCbCrForwardConverter420{TPixel}.cs
  8. 122
      src/ImageSharp/Formats/Jpeg/Components/Encoder/YCbCrForwardConverter444{TPixel}.cs
  9. 61
      src/ImageSharp/Formats/Jpeg/Components/Encoder/YCbCrForwardConverter{TPixel}.cs
  10. 23
      src/ImageSharp/Formats/Jpeg/JpegEncoderCore.cs
  11. 272
      tests/ImageSharp.Tests/Formats/Jpg/RgbToYCbCrConverterTests.cs

35
src/ImageSharp/Formats/Jpeg/Components/Encoder/HuffIndex.cs

@ -1,35 +0,0 @@
// Copyright (c) Six Labors.
// Licensed under the Apache License, Version 2.0.
namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
{
/// <summary>
/// Enumerates the Huffman tables
/// </summary>
internal enum HuffIndex
{
/// <summary>
/// The DC luminance huffman table index
/// </summary>
LuminanceDC = 0,
// ReSharper disable UnusedMember.Local
/// <summary>
/// The AC luminance huffman table index
/// </summary>
LuminanceAC = 1,
/// <summary>
/// The DC chrominance huffman table index
/// </summary>
ChrominanceDC = 2,
/// <summary>
/// The AC chrominance huffman table index
/// </summary>
ChrominanceAC = 3,
// ReSharper restore UnusedMember.Local
}
}

318
src/ImageSharp/Formats/Jpeg/Components/Encoder/HuffmanScanEncoder.cs

@ -7,7 +7,6 @@ using System.Numerics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
using System.Threading;
using SixLabors.ImageSharp.Memory;
using SixLabors.ImageSharp.PixelFormats;
namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
@ -98,8 +97,6 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
private int emitWriteIndex;
private Block8x8 tempBlock;
/// <summary>
/// The output stream. All attempted writes after the first error become no-ops.
/// </summary>
@ -246,321 +243,6 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
}
}
private HuffmanLut[] huffmanTables;
/// <summary>
/// Encodes the image with no subsampling.
/// </summary>
/// <typeparam name="TPixel">The pixel format.</typeparam>
/// <param name="pixels">The pixel accessor providing access to the image pixels.</param>
/// <param name="luminanceQuantTable">Luminance quantization table provided by the callee.</param>
/// <param name="chrominanceQuantTable">Chrominance quantization table provided by the callee.</param>
/// <param name="cancellationToken">The token to monitor for cancellation.</param>
public void Encode444<TPixel>(Image<TPixel> pixels, ref Block8x8F luminanceQuantTable, ref Block8x8F chrominanceQuantTable, CancellationToken cancellationToken)
where TPixel : unmanaged, IPixel<TPixel>
{
FastFloatingPointDCT.AdjustToFDCT(ref luminanceQuantTable);
FastFloatingPointDCT.AdjustToFDCT(ref chrominanceQuantTable);
this.huffmanTables = HuffmanLut.TheHuffmanLut;
// ReSharper disable once InconsistentNaming
int prevDCY = 0, prevDCCb = 0, prevDCCr = 0;
ImageFrame<TPixel> frame = pixels.Frames.RootFrame;
Buffer2D<TPixel> pixelBuffer = frame.PixelBuffer;
RowOctet<TPixel> currentRows = default;
var pixelConverter = new YCbCrForwardConverter444<TPixel>(frame);
for (int y = 0; y < pixels.Height; y += 8)
{
cancellationToken.ThrowIfCancellationRequested();
currentRows.Update(pixelBuffer, y);
for (int x = 0; x < pixels.Width; x += 8)
{
pixelConverter.Convert(x, y, ref currentRows);
prevDCY = this.WriteBlock(
QuantIndex.Luminance,
prevDCY,
ref pixelConverter.Y,
ref luminanceQuantTable);
prevDCCb = this.WriteBlock(
QuantIndex.Chrominance,
prevDCCb,
ref pixelConverter.Cb,
ref chrominanceQuantTable);
prevDCCr = this.WriteBlock(
QuantIndex.Chrominance,
prevDCCr,
ref pixelConverter.Cr,
ref chrominanceQuantTable);
if (this.IsStreamFlushNeeded)
{
this.FlushToStream();
}
}
}
this.FlushRemainingBytes();
}
/// <summary>
/// Encodes the image with subsampling. The Cb and Cr components are each subsampled
/// at a factor of 2 both horizontally and vertically.
/// </summary>
/// <typeparam name="TPixel">The pixel format.</typeparam>
/// <param name="pixels">The pixel accessor providing access to the image pixels.</param>
/// <param name="luminanceQuantTable">Luminance quantization table provided by the callee.</param>
/// <param name="chrominanceQuantTable">Chrominance quantization table provided by the callee.</param>
/// <param name="cancellationToken">The token to monitor for cancellation.</param>
public void Encode420<TPixel>(Image<TPixel> pixels, ref Block8x8F luminanceQuantTable, ref Block8x8F chrominanceQuantTable, CancellationToken cancellationToken)
where TPixel : unmanaged, IPixel<TPixel>
{
FastFloatingPointDCT.AdjustToFDCT(ref luminanceQuantTable);
FastFloatingPointDCT.AdjustToFDCT(ref chrominanceQuantTable);
this.huffmanTables = HuffmanLut.TheHuffmanLut;
// ReSharper disable once InconsistentNaming
int prevDCY = 0, prevDCCb = 0, prevDCCr = 0;
ImageFrame<TPixel> frame = pixels.Frames.RootFrame;
Buffer2D<TPixel> pixelBuffer = frame.PixelBuffer;
RowOctet<TPixel> currentRows = default;
var pixelConverter = new YCbCrForwardConverter420<TPixel>(frame);
for (int y = 0; y < pixels.Height; y += 16)
{
cancellationToken.ThrowIfCancellationRequested();
for (int x = 0; x < pixels.Width; x += 16)
{
for (int i = 0; i < 2; i++)
{
int yOff = i * 8;
currentRows.Update(pixelBuffer, y + yOff);
pixelConverter.Convert(x, y, ref currentRows, i);
prevDCY = this.WriteBlock(
QuantIndex.Luminance,
prevDCY,
ref pixelConverter.YLeft,
ref luminanceQuantTable);
prevDCY = this.WriteBlock(
QuantIndex.Luminance,
prevDCY,
ref pixelConverter.YRight,
ref luminanceQuantTable);
}
prevDCCb = this.WriteBlock(
QuantIndex.Chrominance,
prevDCCb,
ref pixelConverter.Cb,
ref chrominanceQuantTable);
prevDCCr = this.WriteBlock(
QuantIndex.Chrominance,
prevDCCr,
ref pixelConverter.Cr,
ref chrominanceQuantTable);
if (this.IsStreamFlushNeeded)
{
this.FlushToStream();
}
}
}
this.FlushRemainingBytes();
}
/// <summary>
/// Encodes the image with no chroma, just luminance.
/// </summary>
/// <typeparam name="TPixel">The pixel format.</typeparam>
/// <param name="pixels">The pixel accessor providing access to the image pixels.</param>
/// <param name="luminanceQuantTable">Luminance quantization table provided by the callee.</param>
/// <param name="cancellationToken">The token to monitor for cancellation.</param>
public void EncodeGrayscale<TPixel>(Image<TPixel> pixels, ref Block8x8F luminanceQuantTable, CancellationToken cancellationToken)
where TPixel : unmanaged, IPixel<TPixel>
{
FastFloatingPointDCT.AdjustToFDCT(ref luminanceQuantTable);
this.huffmanTables = HuffmanLut.TheHuffmanLut;
// ReSharper disable once InconsistentNaming
int prevDCY = 0;
ImageFrame<TPixel> frame = pixels.Frames.RootFrame;
Buffer2D<TPixel> pixelBuffer = frame.PixelBuffer;
RowOctet<TPixel> currentRows = default;
var pixelConverter = new LuminanceForwardConverter<TPixel>(frame);
for (int y = 0; y < pixels.Height; y += 8)
{
cancellationToken.ThrowIfCancellationRequested();
currentRows.Update(pixelBuffer, y);
for (int x = 0; x < pixels.Width; x += 8)
{
pixelConverter.Convert(x, y, ref currentRows);
prevDCY = this.WriteBlock(
QuantIndex.Luminance,
prevDCY,
ref pixelConverter.Y,
ref luminanceQuantTable);
if (this.IsStreamFlushNeeded)
{
this.FlushToStream();
}
}
}
this.FlushRemainingBytes();
}
/// <summary>
/// Encodes the image with no subsampling and keeps the pixel data as Rgb24.
/// </summary>
/// <typeparam name="TPixel">The pixel format.</typeparam>
/// <param name="pixels">The pixel accessor providing access to the image pixels.</param>
/// <param name="quantTable">Quantization table provided by the callee.</param>
/// <param name="cancellationToken">The token to monitor for cancellation.</param>
public void EncodeRgb<TPixel>(Image<TPixel> pixels, ref Block8x8F quantTable, CancellationToken cancellationToken)
where TPixel : unmanaged, IPixel<TPixel>
{
FastFloatingPointDCT.AdjustToFDCT(ref quantTable);
this.huffmanTables = HuffmanLut.TheHuffmanLut;
// ReSharper disable once InconsistentNaming
int prevDCR = 0, prevDCG = 0, prevDCB = 0;
ImageFrame<TPixel> frame = pixels.Frames.RootFrame;
Buffer2D<TPixel> pixelBuffer = frame.PixelBuffer;
RowOctet<TPixel> currentRows = default;
var pixelConverter = new RgbForwardConverter<TPixel>(frame);
for (int y = 0; y < pixels.Height; y += 8)
{
cancellationToken.ThrowIfCancellationRequested();
currentRows.Update(pixelBuffer, y);
for (int x = 0; x < pixels.Width; x += 8)
{
pixelConverter.Convert(x, y, ref currentRows);
prevDCR = this.WriteBlock(
QuantIndex.Luminance,
prevDCR,
ref pixelConverter.R,
ref quantTable);
prevDCG = this.WriteBlock(
QuantIndex.Luminance,
prevDCG,
ref pixelConverter.G,
ref quantTable);
prevDCB = this.WriteBlock(
QuantIndex.Luminance,
prevDCB,
ref pixelConverter.B,
ref quantTable);
if (this.IsStreamFlushNeeded)
{
this.FlushToStream();
}
}
}
this.FlushRemainingBytes();
}
/// <summary>
/// Writes a block of pixel data using the given quantization table,
/// returning the post-quantized DC value of the DCT-transformed block.
/// The block is in natural (not zig-zag) order.
/// </summary>
/// <param name="index">The quantization table index.</param>
/// <param name="prevDC">The previous DC value.</param>
/// <param name="block">Source block.</param>
/// <param name="quant">Quantization table.</param>
/// <returns>The <see cref="int"/>.</returns>
private int WriteBlock(
QuantIndex index,
int prevDC,
ref Block8x8F block,
ref Block8x8F quant)
{
ref Block8x8 spectralBlock = ref this.tempBlock;
// Shifting level from 0..255 to -128..127
block.AddInPlace(-128f);
// Discrete cosine transform
FastFloatingPointDCT.TransformFDCT(ref block);
// Quantization
Block8x8F.Quantize(ref block, ref spectralBlock, ref quant);
// Emit the DC delta.
int dc = spectralBlock[0];
this.EmitHuffRLE(this.huffmanTables[2 * (int)index].Values, 0, dc - prevDC);
// Emit the AC components.
int[] acHuffTable = this.huffmanTables[(2 * (int)index) + 1].Values;
nint lastValuableIndex = spectralBlock.GetLastNonZeroIndex();
int runLength = 0;
ref short blockRef = ref Unsafe.As<Block8x8, short>(ref spectralBlock);
for (nint zig = 1; zig <= lastValuableIndex; zig++)
{
const int zeroRun1 = 1 << 4;
const int zeroRun16 = 16 << 4;
int ac = Unsafe.Add(ref blockRef, zig);
if (ac == 0)
{
runLength += zeroRun1;
}
else
{
while (runLength >= zeroRun16)
{
this.EmitHuff(acHuffTable, 0xf0);
runLength -= zeroRun16;
}
this.EmitHuffRLE(acHuffTable, runLength, ac);
runLength = 0;
}
}
// if mcu block contains trailing zeros - we must write end of block (EOB) value indicating that current block is over
// this can be done for any number of trailing zeros, even when all 63 ac values are zero
// (Block8x8F.Size - 1) == 63 - last index of the mcu elements
if (lastValuableIndex != Block8x8F.Size - 1)
{
this.EmitHuff(acHuffTable, 0x00);
}
return dc;
}
private void WriteBlock(
JpegComponent component,
ref Block8x8 block,

127
src/ImageSharp/Formats/Jpeg/Components/Encoder/LuminanceForwardConverter{TPixel}.cs

@ -1,127 +0,0 @@
// Copyright (c) Six Labors.
// Licensed under the Apache License, Version 2.0.
using System;
using System.Diagnostics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
#if SUPPORTS_RUNTIME_INTRINSICS
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
#endif
using SixLabors.ImageSharp.Advanced;
using SixLabors.ImageSharp.PixelFormats;
namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
{
/// <summary>
/// On-stack worker struct to efficiently encapsulate the TPixel -> L8 -> Y conversion chain of 8x8 pixel blocks.
/// </summary>
/// <typeparam name="TPixel">The pixel type to work on</typeparam>
internal ref struct LuminanceForwardConverter<TPixel>
where TPixel : unmanaged, IPixel<TPixel>
{
/// <summary>
/// Number of pixels processed per single <see cref="Convert(int, int, ref RowOctet{TPixel})"/> call
/// </summary>
private const int PixelsPerSample = 8 * 8;
/// <summary>
/// The Y component
/// </summary>
public Block8x8F Y;
/// <summary>
/// Temporal 64-pixel span to hold unconverted TPixel data.
/// </summary>
private readonly Span<TPixel> pixelSpan;
/// <summary>
/// Temporal 64-byte span to hold converted <see cref="L8"/> data.
/// </summary>
private readonly Span<L8> l8Span;
/// <summary>
/// Sampled pixel buffer size.
/// </summary>
private readonly Size samplingAreaSize;
/// <summary>
/// <see cref="Configuration"/> for internal operations.
/// </summary>
private readonly Configuration config;
public LuminanceForwardConverter(ImageFrame<TPixel> frame)
{
this.Y = default;
this.pixelSpan = new TPixel[PixelsPerSample].AsSpan();
this.l8Span = new L8[PixelsPerSample].AsSpan();
this.samplingAreaSize = new Size(frame.Width, frame.Height);
this.config = frame.GetConfiguration();
}
/// <summary>
/// Gets size of sampling area from given frame pixel buffer.
/// </summary>
private static Size SampleSize => new(8, 8);
/// <summary>
/// Converts a 8x8 image area inside 'pixels' at position (x,y) placing the result members of the structure (<see cref="Y"/>)
/// </summary>
public void Convert(int x, int y, ref RowOctet<TPixel> currentRows)
{
YCbCrForwardConverter<TPixel>.LoadAndStretchEdges(currentRows, this.pixelSpan, new Point(x, y), SampleSize, this.samplingAreaSize);
PixelOperations<TPixel>.Instance.ToL8(this.config, this.pixelSpan, this.l8Span);
ref Block8x8F yBlock = ref this.Y;
ref L8 l8Start = ref MemoryMarshal.GetReference(this.l8Span);
if (RgbToYCbCrConverterVectorized.IsSupported)
{
ConvertAvx(ref l8Start, ref yBlock);
}
else
{
ConvertScalar(ref l8Start, ref yBlock);
}
}
/// <summary>
/// Converts 8x8 L8 pixel matrix to 8x8 Block of floats using Avx2 Intrinsics.
/// </summary>
/// <param name="l8Start">Start of span of L8 pixels with size of 64</param>
/// <param name="yBlock">8x8 destination matrix of Luminance(Y) converted data</param>
private static void ConvertAvx(ref L8 l8Start, ref Block8x8F yBlock)
{
Debug.Assert(RgbToYCbCrConverterVectorized.IsSupported, "AVX2 is required to run this converter");
#if SUPPORTS_RUNTIME_INTRINSICS
ref Vector128<byte> l8ByteSpan = ref Unsafe.As<L8, Vector128<byte>>(ref l8Start);
ref Vector256<float> destRef = ref yBlock.V0;
const int bytesPerL8Stride = 8;
for (nint i = 0; i < 8; i++)
{
Unsafe.Add(ref destRef, i) = Avx2.ConvertToVector256Single(Avx2.ConvertToVector256Int32(Unsafe.AddByteOffset(ref l8ByteSpan, bytesPerL8Stride * i)));
}
#endif
}
/// <summary>
/// Converts 8x8 L8 pixel matrix to 8x8 Block of floats.
/// </summary>
/// <param name="l8Start">Start of span of L8 pixels with size of 64</param>
/// <param name="yBlock">8x8 destination matrix of Luminance(Y) converted data</param>
private static void ConvertScalar(ref L8 l8Start, ref Block8x8F yBlock)
{
for (int i = 0; i < Block8x8F.Size; i++)
{
ref L8 c = ref Unsafe.Add(ref l8Start, i);
yBlock[i] = c.PackedValue;
}
}
}
}

165
src/ImageSharp/Formats/Jpeg/Components/Encoder/RgbForwardConverter{TPixel}.cs

@ -1,165 +0,0 @@
// Copyright (c) Six Labors.
// Licensed under the Apache License, Version 2.0.
using System;
using System.Diagnostics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
#if SUPPORTS_RUNTIME_INTRINSICS
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
#endif
using SixLabors.ImageSharp.Advanced;
using SixLabors.ImageSharp.PixelFormats;
namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
{
/// <summary>
/// On-stack worker struct to convert TPixel -> Rgb24 of 8x8 pixel blocks.
/// </summary>
/// <typeparam name="TPixel">The pixel type to work on.</typeparam>
internal ref struct RgbForwardConverter<TPixel>
where TPixel : unmanaged, IPixel<TPixel>
{
/// <summary>
/// Number of pixels processed per single <see cref="Convert(int, int, ref RowOctet{TPixel})"/> call
/// </summary>
private const int PixelsPerSample = 8 * 8;
/// <summary>
/// Total byte size of processed pixels converted from TPixel to <see cref="Rgb24"/>
/// </summary>
private const int RgbSpanByteSize = PixelsPerSample * 3;
/// <summary>
/// The Red component.
/// </summary>
public Block8x8F R;
/// <summary>
/// The Green component.
/// </summary>
public Block8x8F G;
/// <summary>
/// The Blue component.
/// </summary>
public Block8x8F B;
/// <summary>
/// Temporal 64-byte span to hold unconverted TPixel data.
/// </summary>
private readonly Span<TPixel> pixelSpan;
/// <summary>
/// Temporal 64-byte span to hold converted Rgb24 data.
/// </summary>
private readonly Span<Rgb24> rgbSpan;
/// <summary>
/// Sampled pixel buffer size.
/// </summary>
private readonly Size samplingAreaSize;
/// <summary>
/// <see cref="Configuration"/> for internal operations.
/// </summary>
private readonly Configuration config;
public RgbForwardConverter(ImageFrame<TPixel> frame)
{
this.R = default;
this.G = default;
this.B = default;
// temporal pixel buffers
this.pixelSpan = new TPixel[PixelsPerSample].AsSpan();
this.rgbSpan = MemoryMarshal.Cast<byte, Rgb24>(new byte[RgbSpanByteSize + RgbToYCbCrConverterVectorized.AvxCompatibilityPadding].AsSpan());
// frame data
this.samplingAreaSize = new Size(frame.Width, frame.Height);
this.config = frame.GetConfiguration();
}
/// <summary>
/// Gets size of sampling area from given frame pixel buffer.
/// </summary>
private static Size SampleSize => new(8, 8);
/// <summary>
/// Converts a 8x8 image area inside 'pixels' at position (x, y) to Rgb24.
/// </summary>
public void Convert(int x, int y, ref RowOctet<TPixel> currentRows)
{
YCbCrForwardConverter<TPixel>.LoadAndStretchEdges(currentRows, this.pixelSpan, new Point(x, y), SampleSize, this.samplingAreaSize);
PixelOperations<TPixel>.Instance.ToRgb24(this.config, this.pixelSpan, this.rgbSpan);
ref Block8x8F redBlock = ref this.R;
ref Block8x8F greenBlock = ref this.G;
ref Block8x8F blueBlock = ref this.B;
if (RgbToYCbCrConverterVectorized.IsSupported)
{
ConvertAvx(this.rgbSpan, ref redBlock, ref greenBlock, ref blueBlock);
}
else
{
ConvertScalar(this.rgbSpan, ref redBlock, ref greenBlock, ref blueBlock);
}
}
/// <summary>
/// Converts 8x8 RGB24 pixel matrix to 8x8 Block of floats using Avx2 Intrinsics.
/// </summary>
/// <param name="rgbSpan">Span of Rgb24 pixels with size of 64</param>
/// <param name="rBlock">8x8 destination matrix of Red converted data</param>
/// <param name="gBlock">8x8 destination matrix of Blue converted data</param>
/// <param name="bBlock">8x8 destination matrix of Green converted data</param>
private static void ConvertAvx(Span<Rgb24> rgbSpan, ref Block8x8F rBlock, ref Block8x8F gBlock, ref Block8x8F bBlock)
{
Debug.Assert(RgbToYCbCrConverterVectorized.IsSupported, "AVX2 is required to run this converter");
#if SUPPORTS_RUNTIME_INTRINSICS
ref Vector256<byte> rgbByteSpan = ref Unsafe.As<Rgb24, Vector256<byte>>(ref MemoryMarshal.GetReference(rgbSpan));
ref Vector256<float> redRef = ref rBlock.V0;
ref Vector256<float> greenRef = ref gBlock.V0;
ref Vector256<float> blueRef = ref bBlock.V0;
var zero = Vector256.Create(0).AsByte();
var extractToLanesMask = Unsafe.As<byte, Vector256<uint>>(ref MemoryMarshal.GetReference(RgbToYCbCrConverterVectorized.MoveFirst24BytesToSeparateLanes));
var extractRgbMask = Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(RgbToYCbCrConverterVectorized.ExtractRgb));
Vector256<byte> rgb, rg, bx;
const int bytesPerRgbStride = 24;
for (nint i = 0; i < 8; i++)
{
rgb = Avx2.PermuteVar8x32(Unsafe.AddByteOffset(ref rgbByteSpan, bytesPerRgbStride * i).AsUInt32(), extractToLanesMask).AsByte();
rgb = Avx2.Shuffle(rgb, extractRgbMask);
rg = Avx2.UnpackLow(rgb, zero);
bx = Avx2.UnpackHigh(rgb, zero);
Unsafe.Add(ref redRef, i) = Avx.ConvertToVector256Single(Avx2.UnpackLow(rg, zero).AsInt32());
Unsafe.Add(ref greenRef, i) = Avx.ConvertToVector256Single(Avx2.UnpackHigh(rg, zero).AsInt32());
Unsafe.Add(ref blueRef, i) = Avx.ConvertToVector256Single(Avx2.UnpackLow(bx, zero).AsInt32());
}
#endif
}
private static void ConvertScalar(Span<Rgb24> rgbSpan, ref Block8x8F redBlock, ref Block8x8F greenBlock, ref Block8x8F blueBlock)
{
ref Rgb24 rgbStart = ref MemoryMarshal.GetReference(rgbSpan);
for (int i = 0; i < Block8x8F.Size; i++)
{
Rgb24 c = Unsafe.Add(ref rgbStart, (nint)(uint)i);
redBlock[i] = c.R;
greenBlock[i] = c.G;
blueBlock[i] = c.B;
}
}
}
}

237
src/ImageSharp/Formats/Jpeg/Components/Encoder/RgbToYCbCrConverterLut.cs

@ -1,237 +0,0 @@
// Copyright (c) Six Labors.
// Licensed under the Apache License, Version 2.0.
using System;
using System.Runtime.CompilerServices;
using SixLabors.ImageSharp.PixelFormats;
namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
{
/// <summary>
/// Provides 8-bit lookup tables for converting from Rgb to YCbCr colorspace.
/// Methods to build the tables are based on libjpeg implementation.
/// </summary>
internal unsafe struct RgbToYCbCrConverterLut
{
/// <summary>
/// The red luminance table
/// </summary>
public fixed int YRTable[256];
/// <summary>
/// The green luminance table
/// </summary>
public fixed int YGTable[256];
/// <summary>
/// The blue luminance table
/// </summary>
public fixed int YBTable[256];
/// <summary>
/// The red blue-chrominance table
/// </summary>
public fixed int CbRTable[256];
/// <summary>
/// The green blue-chrominance table
/// </summary>
public fixed int CbGTable[256];
/// <summary>
/// The blue blue-chrominance table
/// B=>Cb and R=>Cr are the same
/// </summary>
public fixed int CbBTable[256];
/// <summary>
/// The green red-chrominance table
/// </summary>
public fixed int CrGTable[256];
/// <summary>
/// The blue red-chrominance table
/// </summary>
public fixed int CrBTable[256];
// Speediest right-shift on some machines and gives us enough accuracy at 4 decimal places.
private const int ScaleBits = 16;
private const int CBCrOffset = 128 << ScaleBits;
private const int Half = 1 << (ScaleBits - 1);
/// <summary>
/// Initializes the YCbCr tables
/// </summary>
/// <returns>The initialized <see cref="RgbToYCbCrConverterLut"/></returns>
public static RgbToYCbCrConverterLut Create()
{
RgbToYCbCrConverterLut tables = default;
for (int i = 0; i <= 255; i++)
{
// The values for the calculations are left scaled up since we must add them together before rounding.
tables.YRTable[i] = Fix(0.299F) * i;
tables.YGTable[i] = Fix(0.587F) * i;
tables.YBTable[i] = (Fix(0.114F) * i) + Half;
tables.CbRTable[i] = (-Fix(0.168735892F)) * i;
tables.CbGTable[i] = (-Fix(0.331264108F)) * i;
// We use a rounding fudge - factor of 0.5 - epsilon for Cb and Cr.
// This ensures that the maximum output will round to 255
// not 256, and thus that we don't have to range-limit.
//
// B=>Cb and R=>Cr tables are the same
tables.CbBTable[i] = (Fix(0.5F) * i) + CBCrOffset + Half - 1;
tables.CrGTable[i] = (-Fix(0.418687589F)) * i;
tables.CrBTable[i] = (-Fix(0.081312411F)) * i;
}
return tables;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private float CalculateY(byte r, byte g, byte b)
{
// float y = (0.299F * r) + (0.587F * g) + (0.114F * b);
return (this.YRTable[r] + this.YGTable[g] + this.YBTable[b]) >> ScaleBits;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private float CalculateCb(byte r, byte g, byte b)
{
// float cb = 128F + ((-0.168736F * r) - (0.331264F * g) + (0.5F * b));
return (this.CbRTable[r] + this.CbGTable[g] + this.CbBTable[b]) >> ScaleBits;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private float CalculateCr(byte r, byte g, byte b)
{
// float cr = 128F + ((0.5F * r) - (0.418688F * g) - (0.081312F * b));
return (this.CbBTable[r] + this.CrGTable[g] + this.CrBTable[b]) >> ScaleBits;
}
/// <summary>
/// Converts Rgb24 pixels into YCbCr color space with 4:4:4 subsampling sampling of luminance and chroma.
/// </summary>
/// <param name="rgbSpan">Span of Rgb24 pixel data</param>
/// <param name="yBlock">Resulting Y values block</param>
/// <param name="cbBlock">Resulting Cb values block</param>
/// <param name="crBlock">Resulting Cr values block</param>
public void Convert444(Span<Rgb24> rgbSpan, ref Block8x8F yBlock, ref Block8x8F cbBlock, ref Block8x8F crBlock)
{
ref Rgb24 rgbStart = ref rgbSpan[0];
for (int i = 0; i < Block8x8F.Size; i++)
{
Rgb24 c = Unsafe.Add(ref rgbStart, i);
yBlock[i] = this.CalculateY(c.R, c.G, c.B);
cbBlock[i] = this.CalculateCb(c.R, c.G, c.B);
crBlock[i] = this.CalculateCr(c.R, c.G, c.B);
}
}
/// <summary>
/// Converts Rgb24 pixels into YCbCr color space with 4:2:0 subsampling of luminance and chroma.
/// </summary>
/// <remarks>Calculates 2 out of 4 luminance blocks and half of chroma blocks. This method must be called twice per 4x 8x8 DCT blocks with different row param.</remarks>
/// <param name="rgbSpan">Span of Rgb24 pixel data</param>
/// <param name="yBlockLeft">First or "left" resulting Y block</param>
/// <param name="yBlockRight">Second or "right" resulting Y block</param>
/// <param name="cbBlock">Resulting Cb values block</param>
/// <param name="crBlock">Resulting Cr values block</param>
/// <param name="row">Row index of the 16x16 block, 0 or 1</param>
public void Convert420(Span<Rgb24> rgbSpan, ref Block8x8F yBlockLeft, ref Block8x8F yBlockRight, ref Block8x8F cbBlock, ref Block8x8F crBlock, int row)
{
DebugGuard.MustBeBetweenOrEqualTo(row, 0, 1, nameof(row));
ref float yBlockLeftRef = ref Unsafe.As<Block8x8F, float>(ref yBlockLeft);
ref float yBlockRightRef = ref Unsafe.As<Block8x8F, float>(ref yBlockRight);
// 0-31 or 32-63
// upper or lower part
int chromaWriteOffset = row * (Block8x8F.Size / 2);
ref float cbBlockRef = ref Unsafe.Add(ref Unsafe.As<Block8x8F, float>(ref cbBlock), chromaWriteOffset);
ref float crBlockRef = ref Unsafe.Add(ref Unsafe.As<Block8x8F, float>(ref crBlock), chromaWriteOffset);
ref Rgb24 rgbStart = ref rgbSpan[0];
for (int i = 0; i < 8; i += 2)
{
int yBlockWriteOffset = i * 8;
ref Rgb24 stride = ref Unsafe.Add(ref rgbStart, i * 16);
int chromaOffset = 8 * (i / 2);
// left
this.ConvertChunk420(
ref stride,
ref Unsafe.Add(ref yBlockLeftRef, yBlockWriteOffset),
ref Unsafe.Add(ref cbBlockRef, chromaOffset),
ref Unsafe.Add(ref crBlockRef, chromaOffset));
// right
this.ConvertChunk420(
ref Unsafe.Add(ref stride, 8),
ref Unsafe.Add(ref yBlockRightRef, yBlockWriteOffset),
ref Unsafe.Add(ref cbBlockRef, chromaOffset + 4),
ref Unsafe.Add(ref crBlockRef, chromaOffset + 4));
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConvertChunk420(ref Rgb24 stride, ref float yBlock, ref float cbBlock, ref float crBlock)
{
// jpeg 8x8 blocks are processed as 16x16 blocks with 16x8 subpasses (this is done for performance reasons)
// each row is 16 pixels wide thus +16 stride reference offset
// resulting luminance (Y`) are sampled at original resolution thus +8 reference offset
for (int k = 0; k < 8; k += 2)
{
ref float yBlockRef = ref Unsafe.Add(ref yBlock, k);
// top row
Rgb24 px0 = Unsafe.Add(ref stride, k);
Rgb24 px1 = Unsafe.Add(ref stride, k + 1);
yBlockRef = this.CalculateY(px0.R, px0.G, px0.B);
Unsafe.Add(ref yBlockRef, 1) = this.CalculateY(px1.R, px1.G, px1.B);
// bottom row
Rgb24 px2 = Unsafe.Add(ref stride, k + 16);
Rgb24 px3 = Unsafe.Add(ref stride, k + 17);
Unsafe.Add(ref yBlockRef, 8) = this.CalculateY(px2.R, px2.G, px2.B);
Unsafe.Add(ref yBlockRef, 9) = this.CalculateY(px3.R, px3.G, px3.B);
// chroma average for 2x2 pixel block
Unsafe.Add(ref cbBlock, k / 2) = this.CalculateAverageCb(px0, px1, px2, px3);
Unsafe.Add(ref crBlock, k / 2) = this.CalculateAverageCr(px0, px1, px2, px3);
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private float CalculateAverageCb(Rgb24 px0, Rgb24 px1, Rgb24 px2, Rgb24 px3)
{
return 0.25f
* (this.CalculateCb(px0.R, px0.G, px0.B)
+ this.CalculateCb(px1.R, px1.G, px1.B)
+ this.CalculateCb(px2.R, px2.G, px2.B)
+ this.CalculateCb(px3.R, px3.G, px3.B));
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private float CalculateAverageCr(Rgb24 px0, Rgb24 px1, Rgb24 px2, Rgb24 px3)
{
return 0.25f
* (this.CalculateCr(px0.R, px0.G, px0.B)
+ this.CalculateCr(px1.R, px1.G, px1.B)
+ this.CalculateCr(px2.R, px2.G, px2.B)
+ this.CalculateCr(px3.R, px3.G, px3.B));
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static int Fix(float x)
=> (int)((x * (1L << ScaleBits)) + 0.5F);
}
}

259
src/ImageSharp/Formats/Jpeg/Components/Encoder/RgbToYCbCrConverterVectorized.cs

@ -1,259 +0,0 @@
// Copyright (c) Six Labors.
// Licensed under the Apache License, Version 2.0.
using System;
using System.Diagnostics;
#if SUPPORTS_RUNTIME_INTRINSICS
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
#endif
using SixLabors.ImageSharp.PixelFormats;
namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
{
internal static class RgbToYCbCrConverterVectorized
{
public static bool IsSupported
{
get
{
#if SUPPORTS_RUNTIME_INTRINSICS
return Avx2.IsSupported;
#else
return false;
#endif
}
}
public static int AvxCompatibilityPadding
{
// rgb byte matrices contain 8 strides by 8 pixels each, thus 64 pixels total
// Strides are stored sequentially - one big span of 64 * 3 = 192 bytes
// Each stride has exactly 3 * 8 = 24 bytes or 3 * 8 * 8 = 192 bits
// Avx registers are 256 bits so rgb span will be loaded with extra 64 bits from the next stride:
// stride 0 0 - 192 -(+64bits)-> 256
// stride 1 192 - 384 -(+64bits)-> 448
// stride 2 384 - 576 -(+64bits)-> 640
// stride 3 576 - 768 -(+64bits)-> 832
// stride 4 768 - 960 -(+64bits)-> 1024
// stride 5 960 - 1152 -(+64bits)-> 1216
// stride 6 1152 - 1344 -(+64bits)-> 1408
// stride 7 1344 - 1536 -(+64bits)-> 1600 <-- READ ACCESS VIOLATION
//
// Total size of the 64 pixel rgb span: 64 * 3 * 8 = 1536 bits, avx operations require 1600 bits
// This is not permitted - we are reading foreign memory
//
// 8 byte padding to rgb byte span will solve this problem without extra code in converters
get
{
#if SUPPORTS_RUNTIME_INTRINSICS
if (IsSupported)
{
return 8;
}
#endif
return 0;
}
}
#if SUPPORTS_RUNTIME_INTRINSICS
internal static ReadOnlySpan<byte> MoveFirst24BytesToSeparateLanes => new byte[]
{
0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 6, 0, 0, 0,
3, 0, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, 7, 0, 0, 0
};
internal static ReadOnlySpan<byte> ExtractRgb => new byte[]
{
0, 3, 6, 9, 1, 4, 7, 10, 2, 5, 8, 11, 0xFF, 0xFF, 0xFF, 0xFF,
0, 3, 6, 9, 1, 4, 7, 10, 2, 5, 8, 11, 0xFF, 0xFF, 0xFF, 0xFF
};
#endif
/// <summary>
/// Converts 8x8 Rgb24 pixel matrix to YCbCr pixel matrices with 4:4:4 subsampling
/// </summary>
/// <remarks>Total size of rgb span must be 200 bytes</remarks>
/// <param name="rgbSpan">Span of rgb pixels with size of 64</param>
/// <param name="yBlock">8x8 destination matrix of Luminance(Y) converted data</param>
/// <param name="cbBlock">8x8 destination matrix of Chrominance(Cb) converted data</param>
/// <param name="crBlock">8x8 destination matrix of Chrominance(Cr) converted data</param>
public static void Convert444(ReadOnlySpan<Rgb24> rgbSpan, ref Block8x8F yBlock, ref Block8x8F cbBlock, ref Block8x8F crBlock)
{
Debug.Assert(IsSupported, "AVX2 is required to run this converter");
#if SUPPORTS_RUNTIME_INTRINSICS
var f0299 = Vector256.Create(0.299f);
var f0587 = Vector256.Create(0.587f);
var f0114 = Vector256.Create(0.114f);
var fn0168736 = Vector256.Create(-0.168736f);
var fn0331264 = Vector256.Create(-0.331264f);
var f128 = Vector256.Create(128f);
var fn0418688 = Vector256.Create(-0.418688f);
var fn0081312F = Vector256.Create(-0.081312F);
var f05 = Vector256.Create(0.5f);
var zero = Vector256.Create(0).AsByte();
ref Vector256<byte> rgbByteSpan = ref Unsafe.As<Rgb24, Vector256<byte>>(ref MemoryMarshal.GetReference(rgbSpan));
ref Vector256<float> destYRef = ref yBlock.V0;
ref Vector256<float> destCbRef = ref cbBlock.V0;
ref Vector256<float> destCrRef = ref crBlock.V0;
var extractToLanesMask = Unsafe.As<byte, Vector256<uint>>(ref MemoryMarshal.GetReference(MoveFirst24BytesToSeparateLanes));
var extractRgbMask = Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(ExtractRgb));
Vector256<byte> rgb, rg, bx;
Vector256<float> r, g, b;
const int bytesPerRgbStride = 24;
for (int i = 0; i < 8; i++)
{
rgb = Avx2.PermuteVar8x32(Unsafe.AddByteOffset(ref rgbByteSpan, (IntPtr)(bytesPerRgbStride * i)).AsUInt32(), extractToLanesMask).AsByte();
rgb = Avx2.Shuffle(rgb, extractRgbMask);
rg = Avx2.UnpackLow(rgb, zero);
bx = Avx2.UnpackHigh(rgb, zero);
r = Avx.ConvertToVector256Single(Avx2.UnpackLow(rg, zero).AsInt32());
g = Avx.ConvertToVector256Single(Avx2.UnpackHigh(rg, zero).AsInt32());
b = Avx.ConvertToVector256Single(Avx2.UnpackLow(bx, zero).AsInt32());
// (0.299F * r) + (0.587F * g) + (0.114F * b);
Unsafe.Add(ref destYRef, i) = SimdUtils.HwIntrinsics.MultiplyAdd(SimdUtils.HwIntrinsics.MultiplyAdd(Avx.Multiply(f0114, b), f0587, g), f0299, r);
// 128F + ((-0.168736F * r) - (0.331264F * g) + (0.5F * b))
Unsafe.Add(ref destCbRef, i) = Avx.Add(f128, SimdUtils.HwIntrinsics.MultiplyAdd(SimdUtils.HwIntrinsics.MultiplyAdd(Avx.Multiply(f05, b), fn0331264, g), fn0168736, r));
// 128F + ((0.5F * r) - (0.418688F * g) - (0.081312F * b))
Unsafe.Add(ref destCrRef, i) = Avx.Add(f128, SimdUtils.HwIntrinsics.MultiplyAdd(SimdUtils.HwIntrinsics.MultiplyAdd(Avx.Multiply(fn0081312F, b), fn0418688, g), f05, r));
}
#endif
}
/// <summary>
/// Converts 16x8 Rgb24 pixels matrix to 2 Y 8x8 matrices with 4:2:0 subsampling
/// </summary>
public static void Convert420(ReadOnlySpan<Rgb24> rgbSpan, ref Block8x8F yBlockLeft, ref Block8x8F yBlockRight, ref Block8x8F cbBlock, ref Block8x8F crBlock, int row)
{
Debug.Assert(IsSupported, "AVX2 is required to run this converter");
#if SUPPORTS_RUNTIME_INTRINSICS
var f0299 = Vector256.Create(0.299f);
var f0587 = Vector256.Create(0.587f);
var f0114 = Vector256.Create(0.114f);
var fn0168736 = Vector256.Create(-0.168736f);
var fn0331264 = Vector256.Create(-0.331264f);
var f128 = Vector256.Create(128f);
var fn0418688 = Vector256.Create(-0.418688f);
var fn0081312F = Vector256.Create(-0.081312F);
var f05 = Vector256.Create(0.5f);
var zero = Vector256.Create(0).AsByte();
ref Vector256<byte> rgbByteSpan = ref Unsafe.As<Rgb24, Vector256<byte>>(ref MemoryMarshal.GetReference(rgbSpan));
int destOffset = row * 4;
ref Vector256<float> destCbRef = ref Unsafe.Add(ref Unsafe.As<Block8x8F, Vector256<float>>(ref cbBlock), destOffset);
ref Vector256<float> destCrRef = ref Unsafe.Add(ref Unsafe.As<Block8x8F, Vector256<float>>(ref crBlock), destOffset);
var extractToLanesMask = Unsafe.As<byte, Vector256<uint>>(ref MemoryMarshal.GetReference(MoveFirst24BytesToSeparateLanes));
var extractRgbMask = Unsafe.As<byte, Vector256<byte>>(ref MemoryMarshal.GetReference(ExtractRgb));
Vector256<byte> rgb, rg, bx;
Vector256<float> r, g, b;
Span<Vector256<float>> rDataLanes = stackalloc Vector256<float>[4];
Span<Vector256<float>> gDataLanes = stackalloc Vector256<float>[4];
Span<Vector256<float>> bDataLanes = stackalloc Vector256<float>[4];
const int bytesPerRgbStride = 24;
for (int i = 0; i < 4; i++)
{
// 16x2 => 8x1
// left 8x8 column conversions
for (int j = 0; j < 4; j += 2)
{
rgb = Avx2.PermuteVar8x32(Unsafe.AddByteOffset(ref rgbByteSpan, (IntPtr)(bytesPerRgbStride * ((i * 4) + j))).AsUInt32(), extractToLanesMask).AsByte();
rgb = Avx2.Shuffle(rgb, extractRgbMask);
rg = Avx2.UnpackLow(rgb, zero);
bx = Avx2.UnpackHigh(rgb, zero);
r = Avx.ConvertToVector256Single(Avx2.UnpackLow(rg, zero).AsInt32());
g = Avx.ConvertToVector256Single(Avx2.UnpackHigh(rg, zero).AsInt32());
b = Avx.ConvertToVector256Single(Avx2.UnpackLow(bx, zero).AsInt32());
int yBlockVerticalOffset = (i * 2) + ((j & 2) >> 1);
// (0.299F * r) + (0.587F * g) + (0.114F * b);
Unsafe.Add(ref yBlockLeft.V0, yBlockVerticalOffset) = SimdUtils.HwIntrinsics.MultiplyAdd(SimdUtils.HwIntrinsics.MultiplyAdd(Avx.Multiply(f0114, b), f0587, g), f0299, r);
rDataLanes[j] = r;
gDataLanes[j] = g;
bDataLanes[j] = b;
}
// 16x2 => 8x1
// right 8x8 column conversions
for (int j = 1; j < 4; j += 2)
{
rgb = Avx2.PermuteVar8x32(Unsafe.AddByteOffset(ref rgbByteSpan, (IntPtr)(bytesPerRgbStride * ((i * 4) + j))).AsUInt32(), extractToLanesMask).AsByte();
rgb = Avx2.Shuffle(rgb, extractRgbMask);
rg = Avx2.UnpackLow(rgb, zero);
bx = Avx2.UnpackHigh(rgb, zero);
r = Avx.ConvertToVector256Single(Avx2.UnpackLow(rg, zero).AsInt32());
g = Avx.ConvertToVector256Single(Avx2.UnpackHigh(rg, zero).AsInt32());
b = Avx.ConvertToVector256Single(Avx2.UnpackLow(bx, zero).AsInt32());
int yBlockVerticalOffset = (i * 2) + ((j & 2) >> 1);
// (0.299F * r) + (0.587F * g) + (0.114F * b);
Unsafe.Add(ref yBlockRight.V0, yBlockVerticalOffset) = SimdUtils.HwIntrinsics.MultiplyAdd(SimdUtils.HwIntrinsics.MultiplyAdd(Avx.Multiply(f0114, b), f0587, g), f0299, r);
rDataLanes[j] = r;
gDataLanes[j] = g;
bDataLanes[j] = b;
}
r = Scale16x2_8x1(rDataLanes);
g = Scale16x2_8x1(gDataLanes);
b = Scale16x2_8x1(bDataLanes);
// 128F + ((-0.168736F * r) - (0.331264F * g) + (0.5F * b))
Unsafe.Add(ref destCbRef, i) = Avx.Add(f128, SimdUtils.HwIntrinsics.MultiplyAdd(SimdUtils.HwIntrinsics.MultiplyAdd(Avx.Multiply(f05, b), fn0331264, g), fn0168736, r));
// 128F + ((0.5F * r) - (0.418688F * g) - (0.081312F * b))
Unsafe.Add(ref destCrRef, i) = Avx.Add(f128, SimdUtils.HwIntrinsics.MultiplyAdd(SimdUtils.HwIntrinsics.MultiplyAdd(Avx.Multiply(fn0081312F, b), fn0418688, g), f05, r));
}
#endif
}
#if SUPPORTS_RUNTIME_INTRINSICS
/// <summary>
/// Scales 16x2 matrix to 8x1 using 2x2 average
/// </summary>
/// <param name="v">Input matrix consisting of 4 256bit vectors</param>
/// <returns>256bit vector containing upper and lower scaled parts of the input matrix</returns>
[MethodImpl(MethodImplOptions.AggressiveInlining)]
internal static Vector256<float> Scale16x2_8x1(ReadOnlySpan<Vector256<float>> v)
{
Debug.Assert(Avx2.IsSupported, "AVX2 is required to run this converter");
DebugGuard.IsTrue(v.Length == 4, "Input span must consist of 4 elements");
var f025 = Vector256.Create(0.25f);
Vector256<float> left = Avx.Add(v[0], v[2]);
Vector256<float> right = Avx.Add(v[1], v[3]);
Vector256<float> avg2x2 = Avx.Multiply(Avx.HorizontalAdd(left, right), f025);
return Avx2.Permute4x64(avg2x2.AsDouble(), 0b11_01_10_00).AsSingle();
}
#endif
}
}

121
src/ImageSharp/Formats/Jpeg/Components/Encoder/YCbCrForwardConverter420{TPixel}.cs

@ -1,121 +0,0 @@
// Copyright (c) Six Labors.
// Licensed under the Apache License, Version 2.0.
using System;
using System.Runtime.InteropServices;
using SixLabors.ImageSharp.Advanced;
using SixLabors.ImageSharp.PixelFormats;
namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
{
/// <summary>
/// On-stack worker struct to efficiently encapsulate the TPixel -> Rgb24 -> YCbCr conversion chain of 8x8 pixel blocks.
/// </summary>
/// <typeparam name="TPixel">The pixel type to work on</typeparam>
internal ref struct YCbCrForwardConverter420<TPixel>
where TPixel : unmanaged, IPixel<TPixel>
{
/// <summary>
/// Number of pixels processed per single <see cref="Convert(int, int, ref RowOctet{TPixel}, int)"/> call
/// </summary>
private const int PixelsPerSample = 16 * 8;
/// <summary>
/// Total byte size of processed pixels converted from TPixel to <see cref="Rgb24"/>
/// </summary>
private const int RgbSpanByteSize = PixelsPerSample * 3;
/// <summary>
/// The left Y component
/// </summary>
public Block8x8F YLeft;
/// <summary>
/// The left Y component
/// </summary>
public Block8x8F YRight;
/// <summary>
/// The Cb component
/// </summary>
public Block8x8F Cb;
/// <summary>
/// The Cr component
/// </summary>
public Block8x8F Cr;
/// <summary>
/// The color conversion tables
/// </summary>
private RgbToYCbCrConverterLut colorTables;
/// <summary>
/// Temporal 16x8 block to hold TPixel data
/// </summary>
private readonly Span<TPixel> pixelSpan;
/// <summary>
/// Temporal RGB block
/// </summary>
private readonly Span<Rgb24> rgbSpan;
/// <summary>
/// Sampled pixel buffer size
/// </summary>
private readonly Size samplingAreaSize;
/// <summary>
/// <see cref="Configuration"/> for internal operations
/// </summary>
private readonly Configuration config;
public YCbCrForwardConverter420(ImageFrame<TPixel> frame)
{
// matrices would be filled during convert calls
this.YLeft = default;
this.YRight = default;
this.Cb = default;
this.Cr = default;
// temporal pixel buffers
this.pixelSpan = new TPixel[PixelsPerSample].AsSpan();
this.rgbSpan = MemoryMarshal.Cast<byte, Rgb24>(new byte[RgbSpanByteSize + RgbToYCbCrConverterVectorized.AvxCompatibilityPadding].AsSpan());
// frame data
this.samplingAreaSize = new Size(frame.Width, frame.Height);
this.config = frame.GetConfiguration();
// conversion vector fallback data
if (!RgbToYCbCrConverterVectorized.IsSupported)
{
this.colorTables = RgbToYCbCrConverterLut.Create();
}
else
{
this.colorTables = default;
}
}
/// <summary>
/// Gets size of sampling area from given frame pixel buffer.
/// </summary>
private static Size SampleSize => new(16, 8);
public void Convert(int x, int y, ref RowOctet<TPixel> currentRows, int idx)
{
YCbCrForwardConverter<TPixel>.LoadAndStretchEdges(currentRows, this.pixelSpan, new Point(x, y), SampleSize, this.samplingAreaSize);
PixelOperations<TPixel>.Instance.ToRgb24(this.config, this.pixelSpan, this.rgbSpan);
if (RgbToYCbCrConverterVectorized.IsSupported)
{
RgbToYCbCrConverterVectorized.Convert420(this.rgbSpan, ref this.YLeft, ref this.YRight, ref this.Cb, ref this.Cr, idx);
}
else
{
this.colorTables.Convert420(this.rgbSpan, ref this.YLeft, ref this.YRight, ref this.Cb, ref this.Cr, idx);
}
}
}
}

122
src/ImageSharp/Formats/Jpeg/Components/Encoder/YCbCrForwardConverter444{TPixel}.cs

@ -1,122 +0,0 @@
// Copyright (c) Six Labors.
// Licensed under the Apache License, Version 2.0.
using System;
using System.Runtime.InteropServices;
using SixLabors.ImageSharp.Advanced;
using SixLabors.ImageSharp.PixelFormats;
namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
{
/// <summary>
/// On-stack worker struct to efficiently encapsulate the TPixel -> Rgb24 -> YCbCr conversion chain of 8x8 pixel blocks.
/// </summary>
/// <typeparam name="TPixel">The pixel type to work on</typeparam>
internal ref struct YCbCrForwardConverter444<TPixel>
where TPixel : unmanaged, IPixel<TPixel>
{
/// <summary>
/// Number of pixels processed per single <see cref="Convert(int, int, ref RowOctet{TPixel})"/> call
/// </summary>
private const int PixelsPerSample = 8 * 8;
/// <summary>
/// Total byte size of processed pixels converted from TPixel to <see cref="Rgb24"/>
/// </summary>
private const int RgbSpanByteSize = PixelsPerSample * 3;
/// <summary>
/// The Y component
/// </summary>
public Block8x8F Y;
/// <summary>
/// The Cb component
/// </summary>
public Block8x8F Cb;
/// <summary>
/// The Cr component
/// </summary>
public Block8x8F Cr;
/// <summary>
/// The color conversion tables
/// </summary>
private RgbToYCbCrConverterLut colorTables;
/// <summary>
/// Temporal 64-byte span to hold unconverted TPixel data
/// </summary>
private readonly Span<TPixel> pixelSpan;
/// <summary>
/// Temporal 64-byte span to hold converted Rgb24 data
/// </summary>
private readonly Span<Rgb24> rgbSpan;
/// <summary>
/// Sampled pixel buffer size
/// </summary>
private readonly Size samplingAreaSize;
/// <summary>
/// <see cref="Configuration"/> for internal operations
/// </summary>
private readonly Configuration config;
public YCbCrForwardConverter444(ImageFrame<TPixel> frame)
{
// matrices would be filled during convert calls
this.Y = default;
this.Cb = default;
this.Cr = default;
// temporal pixel buffers
this.pixelSpan = new TPixel[PixelsPerSample].AsSpan();
this.rgbSpan = MemoryMarshal.Cast<byte, Rgb24>(new byte[RgbSpanByteSize + RgbToYCbCrConverterVectorized.AvxCompatibilityPadding].AsSpan());
// frame data
this.samplingAreaSize = new Size(frame.Width, frame.Height);
this.config = frame.GetConfiguration();
// conversion vector fallback data
if (!RgbToYCbCrConverterVectorized.IsSupported)
{
this.colorTables = RgbToYCbCrConverterLut.Create();
}
else
{
this.colorTables = default;
}
}
/// <summary>
/// Gets size of sampling area from given frame pixel buffer.
/// </summary>
private static Size SampleSize => new(8, 8);
/// <summary>
/// Converts a 8x8 image area inside 'pixels' at position (x,y) placing the result members of the structure (<see cref="Y"/>, <see cref="Cb"/>, <see cref="Cr"/>)
/// </summary>
public void Convert(int x, int y, ref RowOctet<TPixel> currentRows)
{
YCbCrForwardConverter<TPixel>.LoadAndStretchEdges(currentRows, this.pixelSpan, new Point(x, y), SampleSize, this.samplingAreaSize);
PixelOperations<TPixel>.Instance.ToRgb24(this.config, this.pixelSpan, this.rgbSpan);
ref Block8x8F yBlock = ref this.Y;
ref Block8x8F cbBlock = ref this.Cb;
ref Block8x8F crBlock = ref this.Cr;
if (RgbToYCbCrConverterVectorized.IsSupported)
{
RgbToYCbCrConverterVectorized.Convert444(this.rgbSpan, ref yBlock, ref cbBlock, ref crBlock);
}
else
{
this.colorTables.Convert444(this.rgbSpan, ref yBlock, ref cbBlock, ref crBlock);
}
}
}
}

61
src/ImageSharp/Formats/Jpeg/Components/Encoder/YCbCrForwardConverter{TPixel}.cs

@ -1,61 +0,0 @@
// Copyright (c) Six Labors.
// Licensed under the Apache License, Version 2.0.
using System;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
using SixLabors.ImageSharp.PixelFormats;
namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder
{
internal static class YCbCrForwardConverter<TPixel>
where TPixel : unmanaged, IPixel<TPixel>
{
public static void LoadAndStretchEdges(RowOctet<TPixel> source, Span<TPixel> dest, Point start, Size sampleSize, Size totalSize)
{
DebugGuard.MustBeBetweenOrEqualTo(start.X, 0, totalSize.Width - 1, nameof(start.X));
DebugGuard.MustBeBetweenOrEqualTo(start.Y, 0, totalSize.Height - 1, nameof(start.Y));
int width = Math.Min(sampleSize.Width, totalSize.Width - start.X);
int height = Math.Min(sampleSize.Height, totalSize.Height - start.Y);
uint byteWidth = (uint)(width * Unsafe.SizeOf<TPixel>());
int remainderXCount = sampleSize.Width - width;
ref byte blockStart = ref MemoryMarshal.GetReference(MemoryMarshal.Cast<TPixel, byte>(dest));
int rowSizeInBytes = sampleSize.Width * Unsafe.SizeOf<TPixel>();
for (int y = 0; y < height; y++)
{
Span<TPixel> row = source[y];
ref byte s = ref Unsafe.As<TPixel, byte>(ref row[start.X]);
ref byte d = ref Unsafe.Add(ref blockStart, y * rowSizeInBytes);
Unsafe.CopyBlock(ref d, ref s, byteWidth);
ref TPixel last = ref Unsafe.Add(ref Unsafe.As<byte, TPixel>(ref d), width - 1);
for (int x = 1; x <= remainderXCount; x++)
{
Unsafe.Add(ref last, x) = last;
}
}
int remainderYCount = sampleSize.Height - height;
if (remainderYCount == 0)
{
return;
}
ref byte lastRowStart = ref Unsafe.Add(ref blockStart, (height - 1) * rowSizeInBytes);
for (int y = 1; y <= remainderYCount; y++)
{
ref byte remStart = ref Unsafe.Add(ref lastRowStart, rowSizeInBytes * y);
Unsafe.CopyBlock(ref remStart, ref lastRowStart, (uint)rowSizeInBytes);
}
}
}
}

23
src/ImageSharp/Formats/Jpeg/JpegEncoderCore.cs

@ -90,7 +90,7 @@ namespace SixLabors.ImageSharp.Formats.Jpeg
cancellationToken.ThrowIfCancellationRequested();
var frame = new JpegFrame(this.frameConfig, Configuration.Default.MemoryAllocator, image, GetTargetColorSpace(this.frameConfig.EncodingColor));
var frame = new JpegFrame(this.frameConfig, Configuration.Default.MemoryAllocator, image, this.frameConfig.ColorType);
this.scanEncoder = new HuffmanScanEncoder(frame.BlocksPerMcu, stream);
this.outputStream = stream;
@ -142,27 +142,6 @@ namespace SixLabors.ImageSharp.Formats.Jpeg
this.WriteEndOfImageMarker();
stream.Flush();
static JpegColorSpace GetTargetColorSpace(JpegEncodingColor colorType)
{
switch (colorType)
{
case JpegEncodingColor.YCbCrRatio444:
case JpegEncodingColor.YCbCrRatio422:
case JpegEncodingColor.YCbCrRatio420:
case JpegEncodingColor.YCbCrRatio411:
case JpegEncodingColor.YCbCrRatio410:
return JpegColorSpace.YCbCr;
case JpegEncodingColor.Rgb:
return JpegColorSpace.RGB;
case JpegEncodingColor.Cmyk:
return JpegColorSpace.Cmyk;
case JpegEncodingColor.Luminance:
return JpegColorSpace.Grayscale;
default:
throw new NotImplementedException($"Unknown output color space: {colorType}");
}
}
}
/// <summary>

272
tests/ImageSharp.Tests/Formats/Jpg/RgbToYCbCrConverterTests.cs

@ -1,272 +0,0 @@
// Copyright (c) Six Labors.
// Licensed under the Apache License, Version 2.0.
using System;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
#if SUPPORTS_RUNTIME_INTRINSICS
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.X86;
#endif
using SixLabors.ImageSharp.ColorSpaces;
using SixLabors.ImageSharp.Formats.Jpeg.Components;
using SixLabors.ImageSharp.Formats.Jpeg.Components.Encoder;
using SixLabors.ImageSharp.PixelFormats;
using SixLabors.ImageSharp.Tests.Colorspaces.Conversion;
using Xunit;
using Xunit.Abstractions;
// ReSharper disable InconsistentNaming
namespace SixLabors.ImageSharp.Tests.Formats.Jpg
{
public class RgbToYCbCrConverterTests
{
public RgbToYCbCrConverterTests(ITestOutputHelper output)
{
this.Output = output;
}
private ITestOutputHelper Output { get; }
[Fact]
public void TestConverterLut444()
{
int dataSize = 8 * 8;
Rgb24[] data = CreateTestData(dataSize);
var target = RgbToYCbCrConverterLut.Create();
Block8x8F y = default;
Block8x8F cb = default;
Block8x8F cr = default;
target.Convert444(data.AsSpan(), ref y, ref cb, ref cr);
Verify444(data, ref y, ref cb, ref cr, new ApproximateColorSpaceComparer(1F));
}
[Fact]
public void TestConverterVectorized444()
{
if (!RgbToYCbCrConverterVectorized.IsSupported)
{
this.Output.WriteLine("No AVX and/or FMA present, skipping test!");
return;
}
int dataSize = 8 * 8;
Rgb24[] data = CreateTestData(dataSize);
Block8x8F y = default;
Block8x8F cb = default;
Block8x8F cr = default;
RgbToYCbCrConverterVectorized.Convert444(data.AsSpan(), ref y, ref cb, ref cr);
Verify444(data, ref y, ref cb, ref cr, new ApproximateColorSpaceComparer(0.0001F));
}
[Fact]
public void TestConverterLut420()
{
int dataSize = 16 * 16;
Span<Rgb24> data = CreateTestData(dataSize).AsSpan();
var target = RgbToYCbCrConverterLut.Create();
var yBlocks = new Block8x8F[4];
var cb = default(Block8x8F);
var cr = default(Block8x8F);
target.Convert420(data, ref yBlocks[0], ref yBlocks[1], ref cb, ref cr, 0);
target.Convert420(data.Slice(16 * 8), ref yBlocks[2], ref yBlocks[3], ref cb, ref cr, 1);
Verify420(data, yBlocks, ref cb, ref cr, new ApproximateFloatComparer(1F));
}
[Fact]
public void TestConverterVectorized420()
{
if (!RgbToYCbCrConverterVectorized.IsSupported)
{
this.Output.WriteLine("No AVX and/or FMA present, skipping test!");
return;
}
int dataSize = 16 * 16;
Span<Rgb24> data = CreateTestData(dataSize).AsSpan();
var yBlocks = new Block8x8F[4];
var cb = default(Block8x8F);
var cr = default(Block8x8F);
RgbToYCbCrConverterVectorized.Convert420(data, ref yBlocks[0], ref yBlocks[1], ref cb, ref cr, 0);
RgbToYCbCrConverterVectorized.Convert420(data.Slice(16 * 8), ref yBlocks[2], ref yBlocks[3], ref cb, ref cr, 1);
Verify420(data, yBlocks, ref cb, ref cr, new ApproximateFloatComparer(1F));
}
#if SUPPORTS_RUNTIME_INTRINSICS
[Theory]
[InlineData(1)]
[InlineData(2)]
[InlineData(3)]
public void Scale16x2_8x1(int seed)
{
if (!Avx2.IsSupported)
{
return;
}
Span<float> data = new Random(seed).GenerateRandomFloatArray(Vector256<float>.Count * 4, -1000, 1000);
// Act:
Vector256<float> resultVector = RgbToYCbCrConverterVectorized.Scale16x2_8x1(MemoryMarshal.Cast<float, Vector256<float>>(data));
ref float result = ref Unsafe.As<Vector256<float>, float>(ref resultVector);
// Assert:
// Comparison epsilon is tricky but 10^(-4) is good enough (?)
var comparer = new ApproximateFloatComparer(0.0001f);
for (int i = 0; i < Vector256<float>.Count; i++)
{
float actual = Unsafe.Add(ref result, i);
float expected = CalculateAverage16x2_8x1(data, i);
Assert.True(comparer.Equals(actual, expected), $"Pos {i}, Expected: {expected}, Actual: {actual}");
}
static float CalculateAverage16x2_8x1(Span<float> data, int index)
{
int upIdx = index * 2;
int lowIdx = (index + 8) * 2;
return 0.25f * (data[upIdx] + data[upIdx + 1] + data[lowIdx] + data[lowIdx + 1]);
}
}
#endif
private static void Verify444(
ReadOnlySpan<Rgb24> data,
ref Block8x8F yResult,
ref Block8x8F cbResult,
ref Block8x8F crResult,
ApproximateColorSpaceComparer comparer)
{
Block8x8F y = default;
Block8x8F cb = default;
Block8x8F cr = default;
RgbToYCbCr(data, ref y, ref cb, ref cr);
for (int i = 0; i < Block8x8F.Size; i++)
{
Assert.True(comparer.Equals(new YCbCr(y[i], cb[i], cr[i]), new YCbCr(yResult[i], cbResult[i], crResult[i])), $"Pos {i}, Expected {y[i]} == {yResult[i]}, {cb[i]} == {cbResult[i]}, {cr[i]} == {crResult[i]}");
}
}
private static void Verify420(
ReadOnlySpan<Rgb24> data,
Block8x8F[] yResult,
ref Block8x8F cbResult,
ref Block8x8F crResult,
ApproximateFloatComparer comparer)
{
var trueBlock = default(Block8x8F);
var cbTrue = new Block8x8F[4];
var crTrue = new Block8x8F[4];
Span<Rgb24> tempData = new Rgb24[8 * 8].AsSpan();
// top left
Copy8x8(data, tempData);
RgbToYCbCr(tempData, ref trueBlock, ref cbTrue[0], ref crTrue[0]);
VerifyBlock(ref yResult[0], ref trueBlock, comparer);
// top right
Copy8x8(data.Slice(8), tempData);
RgbToYCbCr(tempData, ref trueBlock, ref cbTrue[1], ref crTrue[1]);
VerifyBlock(ref yResult[1], ref trueBlock, comparer);
// bottom left
Copy8x8(data.Slice(8 * 16), tempData);
RgbToYCbCr(tempData, ref trueBlock, ref cbTrue[2], ref crTrue[2]);
VerifyBlock(ref yResult[2], ref trueBlock, comparer);
// bottom right
Copy8x8(data.Slice((8 * 16) + 8), tempData);
RgbToYCbCr(tempData, ref trueBlock, ref cbTrue[3], ref crTrue[3]);
VerifyBlock(ref yResult[3], ref trueBlock, comparer);
// verify Cb
Scale16X16To8X8(ref trueBlock, cbTrue);
VerifyBlock(ref cbResult, ref trueBlock, comparer);
// verify Cr
Scale16X16To8X8(ref trueBlock, crTrue);
VerifyBlock(ref crResult, ref trueBlock, comparer);
// extracts 8x8 blocks from 16x8 memory region
static void Copy8x8(ReadOnlySpan<Rgb24> source, Span<Rgb24> dest)
{
for (int i = 0; i < 8; i++)
{
source.Slice(i * 16, 8).CopyTo(dest.Slice(i * 8));
}
}
// scales 16x16 to 8x8, used in chroma subsampling tests
static void Scale16X16To8X8(ref Block8x8F dest, ReadOnlySpan<Block8x8F> source)
{
for (int i = 0; i < 4; i++)
{
int dstOff = ((i & 2) << 4) | ((i & 1) << 2);
Block8x8F iSource = source[i];
for (int y = 0; y < 4; y++)
{
for (int x = 0; x < 4; x++)
{
int j = (16 * y) + (2 * x);
float sum = iSource[j] + iSource[j + 1] + iSource[j + 8] + iSource[j + 9];
dest[(8 * y) + x + dstOff] = (sum + 2) * .25F;
}
}
}
}
}
private static void RgbToYCbCr(ReadOnlySpan<Rgb24> data, ref Block8x8F y, ref Block8x8F cb, ref Block8x8F cr)
{
for (int i = 0; i < data.Length; i++)
{
int r = data[i].R;
int g = data[i].G;
int b = data[i].B;
y[i] = (0.299F * r) + (0.587F * g) + (0.114F * b);
cb[i] = 128F + ((-0.168736F * r) - (0.331264F * g) + (0.5F * b));
cr[i] = 128F + ((0.5F * r) - (0.418688F * g) - (0.081312F * b));
}
}
private static void VerifyBlock(ref Block8x8F res, ref Block8x8F target, ApproximateFloatComparer comparer)
{
for (int i = 0; i < Block8x8F.Size; i++)
{
Assert.True(comparer.Equals(res[i], target[i]), $"Pos {i}, Expected: {target[i]}, Got: {res[i]}");
}
}
private static Rgb24[] CreateTestData(int size)
{
var data = new Rgb24[size];
var r = new Random();
var random = new byte[3];
for (int i = 0; i < data.Length; i++)
{
r.NextBytes(random);
data[i] = new Rgb24(random[0], random[1], random[2]);
}
return data;
}
}
}
Loading…
Cancel
Save