|
|
@ -17,6 +17,11 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
/// </summary>
|
|
|
/// </summary>
|
|
|
internal static unsafe class PredictorEncoder |
|
|
internal static unsafe class PredictorEncoder |
|
|
{ |
|
|
{ |
|
|
|
|
|
private static readonly sbyte[][] Offset = |
|
|
|
|
|
{ |
|
|
|
|
|
new sbyte[] { 0, -1 }, new sbyte[] { 0, 1 }, new sbyte[] { -1, 0 }, new sbyte[] { 1, 0 }, new sbyte[] { -1, -1 }, new sbyte[] { -1, 1 }, new sbyte[] { 1, -1 }, new sbyte[] { 1, 1 } |
|
|
|
|
|
}; |
|
|
|
|
|
|
|
|
private const int GreenRedToBlueNumAxis = 8; |
|
|
private const int GreenRedToBlueNumAxis = 8; |
|
|
|
|
|
|
|
|
private const int GreenRedToBlueMaxIters = 7; |
|
|
private const int GreenRedToBlueMaxIters = 7; |
|
|
@ -29,6 +34,25 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
|
|
|
|
|
|
private const int PredLowEffort = 11; |
|
|
private const int PredLowEffort = 11; |
|
|
|
|
|
|
|
|
|
|
|
#if SUPPORTS_RUNTIME_INTRINSICS
|
|
|
|
|
|
private static readonly Vector128<byte> CollectColorRedTransformsGreenMask = Vector128.Create(0x00ff00).AsByte(); |
|
|
|
|
|
|
|
|
|
|
|
private static readonly Vector128<byte> CollectColorRedTransformsAndMask = Vector128.Create((short)0xff).AsByte(); |
|
|
|
|
|
|
|
|
|
|
|
private static readonly Vector128<byte> CollectColorBlueTransformsGreenMask = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255); |
|
|
|
|
|
|
|
|
|
|
|
private static readonly Vector128<byte> CollectColorBlueTransformsGreenBlueMask = Vector128.Create(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); |
|
|
|
|
|
|
|
|
|
|
|
private static readonly Vector128<byte> CollectColorBlueTransformsBlueMask = Vector128.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0); |
|
|
|
|
|
|
|
|
|
|
|
private static readonly Vector128<byte> CollectColorBlueTransformsShuffleLowMask = Vector128.Create(255, 2, 255, 6, 255, 10, 255, 14, 255, 255, 255, 255, 255, 255, 255, 255); |
|
|
|
|
|
|
|
|
|
|
|
private static readonly Vector128<byte> CollectColorBlueTransformsShuffleHighMask = Vector128.Create(255, 255, 255, 255, 255, 255, 255, 255, 255, 2, 255, 6, 255, 10, 255, 14); |
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
|
|
// This uses C#'s compiler optimization to refer to assembly's static data directly.
|
|
|
|
|
|
private static ReadOnlySpan<sbyte> DeltaLut => new sbyte[] { 16, 16, 8, 4, 2, 2, 2 }; |
|
|
|
|
|
|
|
|
/// <summary>
|
|
|
/// <summary>
|
|
|
/// Finds the best predictor for each tile, and converts the image to residuals
|
|
|
/// Finds the best predictor for each tile, and converts the image to residuals
|
|
|
/// with respect to predictions. If nearLosslessQuality < 100, applies
|
|
|
/// with respect to predictions. If nearLosslessQuality < 100, applies
|
|
|
@ -41,6 +65,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
Span<uint> bgra, |
|
|
Span<uint> bgra, |
|
|
Span<uint> bgraScratch, |
|
|
Span<uint> bgraScratch, |
|
|
Span<uint> image, |
|
|
Span<uint> image, |
|
|
|
|
|
int[][] histoArgb, |
|
|
|
|
|
int[][] bestHisto, |
|
|
bool nearLossless, |
|
|
bool nearLossless, |
|
|
int nearLosslessQuality, |
|
|
int nearLosslessQuality, |
|
|
WebpTransparentColorMode transparentColorMode, |
|
|
WebpTransparentColorMode transparentColorMode, |
|
|
@ -81,6 +107,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
histo, |
|
|
histo, |
|
|
bgraScratch, |
|
|
bgraScratch, |
|
|
bgra, |
|
|
bgra, |
|
|
|
|
|
histoArgb, |
|
|
|
|
|
bestHisto, |
|
|
maxQuantization, |
|
|
maxQuantization, |
|
|
transparentColorMode, |
|
|
transparentColorMode, |
|
|
usedSubtractGreen, |
|
|
usedSubtractGreen, |
|
|
@ -107,7 +135,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
lowEffort); |
|
|
lowEffort); |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
public static void ColorSpaceTransform(int width, int height, int bits, int quality, Span<uint> bgra, Span<uint> image) |
|
|
public static void ColorSpaceTransform(int width, int height, int bits, int quality, Span<uint> bgra, Span<uint> image, Span<int> scratch) |
|
|
{ |
|
|
{ |
|
|
int maxTileSize = 1 << bits; |
|
|
int maxTileSize = 1 << bits; |
|
|
int tileXSize = LosslessUtils.SubSampleSize(width, bits); |
|
|
int tileXSize = LosslessUtils.SubSampleSize(width, bits); |
|
|
@ -141,7 +169,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
height, |
|
|
height, |
|
|
accumulatedRedHisto, |
|
|
accumulatedRedHisto, |
|
|
accumulatedBlueHisto, |
|
|
accumulatedBlueHisto, |
|
|
bgra); |
|
|
bgra, |
|
|
|
|
|
scratch); |
|
|
|
|
|
|
|
|
image[offset] = MultipliersToColorCode(prevX); |
|
|
image[offset] = MultipliersToColorCode(prevX); |
|
|
CopyTileWithColorTransform(width, height, tileXOffset, tileYOffset, maxTileSize, prevX, bgra); |
|
|
CopyTileWithColorTransform(width, height, tileXOffset, tileYOffset, maxTileSize, prevX, bgra); |
|
|
@ -190,6 +219,8 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
int[][] accumulated, |
|
|
int[][] accumulated, |
|
|
Span<uint> argbScratch, |
|
|
Span<uint> argbScratch, |
|
|
Span<uint> argb, |
|
|
Span<uint> argb, |
|
|
|
|
|
int[][] histoArgb, |
|
|
|
|
|
int[][] bestHisto, |
|
|
int maxQuantization, |
|
|
int maxQuantization, |
|
|
WebpTransparentColorMode transparentColorMode, |
|
|
WebpTransparentColorMode transparentColorMode, |
|
|
bool usedSubtractGreen, |
|
|
bool usedSubtractGreen, |
|
|
@ -225,21 +256,14 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
float bestDiff = MaxDiffCost; |
|
|
float bestDiff = MaxDiffCost; |
|
|
int bestMode = 0; |
|
|
int bestMode = 0; |
|
|
uint[] residuals = new uint[1 << WebpConstants.MaxTransformBits]; |
|
|
uint[] residuals = new uint[1 << WebpConstants.MaxTransformBits]; |
|
|
int[][] histoArgb = new int[4][]; |
|
|
|
|
|
int[][] bestHisto = new int[4][]; |
|
|
|
|
|
for (int i = 0; i < 4; i++) |
|
|
for (int i = 0; i < 4; i++) |
|
|
{ |
|
|
{ |
|
|
histoArgb[i] = new int[256]; |
|
|
histoArgb[i].AsSpan().Clear(); |
|
|
bestHisto[i] = new int[256]; |
|
|
bestHisto[i].AsSpan().Clear(); |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
for (int mode = 0; mode < numPredModes; mode++) |
|
|
for (int mode = 0; mode < numPredModes; mode++) |
|
|
{ |
|
|
{ |
|
|
for (int i = 0; i < 4; i++) |
|
|
|
|
|
{ |
|
|
|
|
|
histoArgb[i].AsSpan().Fill(0); |
|
|
|
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
if (startY > 0) |
|
|
if (startY > 0) |
|
|
{ |
|
|
{ |
|
|
// Read the row above the tile which will become the first upper_row.
|
|
|
// Read the row above the tile which will become the first upper_row.
|
|
|
@ -303,6 +327,11 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
bestDiff = curDiff; |
|
|
bestDiff = curDiff; |
|
|
bestMode = mode; |
|
|
bestMode = mode; |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
|
|
|
for (int i = 0; i < 4; i++) |
|
|
|
|
|
{ |
|
|
|
|
|
histoArgb[i].AsSpan().Clear(); |
|
|
|
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
for (int i = 0; i < 4; i++) |
|
|
for (int i = 0; i < 4; i++) |
|
|
@ -826,7 +855,19 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
} |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
private static Vp8LMultipliers GetBestColorTransformForTile(int tileX, int tileY, int bits, Vp8LMultipliers prevX, Vp8LMultipliers prevY, int quality, int xSize, int ySize, int[] accumulatedRedHisto, int[] accumulatedBlueHisto, Span<uint> argb) |
|
|
private static Vp8LMultipliers GetBestColorTransformForTile( |
|
|
|
|
|
int tileX, |
|
|
|
|
|
int tileY, |
|
|
|
|
|
int bits, |
|
|
|
|
|
Vp8LMultipliers prevX, |
|
|
|
|
|
Vp8LMultipliers prevY, |
|
|
|
|
|
int quality, |
|
|
|
|
|
int xSize, |
|
|
|
|
|
int ySize, |
|
|
|
|
|
int[] accumulatedRedHisto, |
|
|
|
|
|
int[] accumulatedBlueHisto, |
|
|
|
|
|
Span<uint> argb, |
|
|
|
|
|
Span<int> scratch) |
|
|
{ |
|
|
{ |
|
|
int maxTileSize = 1 << bits; |
|
|
int maxTileSize = 1 << bits; |
|
|
int tileYOffset = tileY * maxTileSize; |
|
|
int tileYOffset = tileY * maxTileSize; |
|
|
@ -839,18 +880,28 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
|
|
|
|
|
|
var bestTx = default(Vp8LMultipliers); |
|
|
var bestTx = default(Vp8LMultipliers); |
|
|
|
|
|
|
|
|
GetBestGreenToRed(tileArgb, xSize, tileWidth, tileHeight, prevX, prevY, quality, accumulatedRedHisto, ref bestTx); |
|
|
GetBestGreenToRed(tileArgb, xSize, scratch, tileWidth, tileHeight, prevX, prevY, quality, accumulatedRedHisto, ref bestTx); |
|
|
|
|
|
|
|
|
GetBestGreenRedToBlue(tileArgb, xSize, tileWidth, tileHeight, prevX, prevY, quality, accumulatedBlueHisto, ref bestTx); |
|
|
GetBestGreenRedToBlue(tileArgb, xSize, scratch, tileWidth, tileHeight, prevX, prevY, quality, accumulatedBlueHisto, ref bestTx); |
|
|
|
|
|
|
|
|
return bestTx; |
|
|
return bestTx; |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
private static void GetBestGreenToRed(Span<uint> argb, int stride, int tileWidth, int tileHeight, Vp8LMultipliers prevX, Vp8LMultipliers prevY, int quality, int[] accumulatedRedHisto, ref Vp8LMultipliers bestTx) |
|
|
private static void GetBestGreenToRed( |
|
|
|
|
|
Span<uint> argb, |
|
|
|
|
|
int stride, |
|
|
|
|
|
Span<int> scratch, |
|
|
|
|
|
int tileWidth, |
|
|
|
|
|
int tileHeight, |
|
|
|
|
|
Vp8LMultipliers prevX, |
|
|
|
|
|
Vp8LMultipliers prevY, |
|
|
|
|
|
int quality, |
|
|
|
|
|
int[] accumulatedRedHisto, |
|
|
|
|
|
ref Vp8LMultipliers bestTx) |
|
|
{ |
|
|
{ |
|
|
int maxIters = 4 + ((7 * quality) >> 8); // in range [4..6]
|
|
|
int maxIters = 4 + ((7 * quality) >> 8); // in range [4..6]
|
|
|
int greenToRedBest = 0; |
|
|
int greenToRedBest = 0; |
|
|
double bestDiff = GetPredictionCostCrossColorRed(argb, stride, tileWidth, tileHeight, prevX, prevY, greenToRedBest, accumulatedRedHisto); |
|
|
double bestDiff = GetPredictionCostCrossColorRed(argb, stride, scratch, tileWidth, tileHeight, prevX, prevY, greenToRedBest, accumulatedRedHisto); |
|
|
for (int iter = 0; iter < maxIters; iter++) |
|
|
for (int iter = 0; iter < maxIters; iter++) |
|
|
{ |
|
|
{ |
|
|
// ColorTransformDelta is a 3.5 bit fixed point, so 32 is equal to
|
|
|
// ColorTransformDelta is a 3.5 bit fixed point, so 32 is equal to
|
|
|
@ -862,7 +913,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
for (int offset = -delta; offset <= delta; offset += 2 * delta) |
|
|
for (int offset = -delta; offset <= delta; offset += 2 * delta) |
|
|
{ |
|
|
{ |
|
|
int greenToRedCur = offset + greenToRedBest; |
|
|
int greenToRedCur = offset + greenToRedBest; |
|
|
double curDiff = GetPredictionCostCrossColorRed(argb, stride, tileWidth, tileHeight, prevX, prevY, greenToRedCur, accumulatedRedHisto); |
|
|
double curDiff = GetPredictionCostCrossColorRed(argb, stride, scratch, tileWidth, tileHeight, prevX, prevY, greenToRedCur, accumulatedRedHisto); |
|
|
if (curDiff < bestDiff) |
|
|
if (curDiff < bestDiff) |
|
|
{ |
|
|
{ |
|
|
bestDiff = curDiff; |
|
|
bestDiff = curDiff; |
|
|
@ -874,24 +925,22 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
bestTx.GreenToRed = (byte)(greenToRedBest & 0xff); |
|
|
bestTx.GreenToRed = (byte)(greenToRedBest & 0xff); |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
private static void GetBestGreenRedToBlue(Span<uint> argb, int stride, int tileWidth, int tileHeight, Vp8LMultipliers prevX, Vp8LMultipliers prevY, int quality, int[] accumulatedBlueHisto, ref Vp8LMultipliers bestTx) |
|
|
private static void GetBestGreenRedToBlue(Span<uint> argb, int stride, Span<int> scratch, int tileWidth, int tileHeight, Vp8LMultipliers prevX, Vp8LMultipliers prevY, int quality, int[] accumulatedBlueHisto, ref Vp8LMultipliers bestTx) |
|
|
{ |
|
|
{ |
|
|
int iters = (quality < 25) ? 1 : (quality > 50) ? GreenRedToBlueMaxIters : 4; |
|
|
int iters = (quality < 25) ? 1 : (quality > 50) ? GreenRedToBlueMaxIters : 4; |
|
|
int greenToBlueBest = 0; |
|
|
int greenToBlueBest = 0; |
|
|
int redToBlueBest = 0; |
|
|
int redToBlueBest = 0; |
|
|
sbyte[][] offset = { new sbyte[] { 0, -1 }, new sbyte[] { 0, 1 }, new sbyte[] { -1, 0 }, new sbyte[] { 1, 0 }, new sbyte[] { -1, -1 }, new sbyte[] { -1, 1 }, new sbyte[] { 1, -1 }, new sbyte[] { 1, 1 } }; |
|
|
|
|
|
sbyte[] deltaLut = { 16, 16, 8, 4, 2, 2, 2 }; |
|
|
|
|
|
|
|
|
|
|
|
// Initial value at origin:
|
|
|
// Initial value at origin:
|
|
|
double bestDiff = GetPredictionCostCrossColorBlue(argb, stride, tileWidth, tileHeight, prevX, prevY, greenToBlueBest, redToBlueBest, accumulatedBlueHisto); |
|
|
double bestDiff = GetPredictionCostCrossColorBlue(argb, stride, scratch, tileWidth, tileHeight, prevX, prevY, greenToBlueBest, redToBlueBest, accumulatedBlueHisto); |
|
|
for (int iter = 0; iter < iters; iter++) |
|
|
for (int iter = 0; iter < iters; iter++) |
|
|
{ |
|
|
{ |
|
|
int delta = deltaLut[iter]; |
|
|
int delta = DeltaLut[iter]; |
|
|
for (int axis = 0; axis < GreenRedToBlueNumAxis; axis++) |
|
|
for (int axis = 0; axis < GreenRedToBlueNumAxis; axis++) |
|
|
{ |
|
|
{ |
|
|
int greenToBlueCur = (offset[axis][0] * delta) + greenToBlueBest; |
|
|
int greenToBlueCur = (Offset[axis][0] * delta) + greenToBlueBest; |
|
|
int redToBlueCur = (offset[axis][1] * delta) + redToBlueBest; |
|
|
int redToBlueCur = (Offset[axis][1] * delta) + redToBlueBest; |
|
|
double curDiff = GetPredictionCostCrossColorBlue(argb, stride, tileWidth, tileHeight, prevX, prevY, greenToBlueCur, redToBlueCur, accumulatedBlueHisto); |
|
|
double curDiff = GetPredictionCostCrossColorBlue(argb, stride, scratch, tileWidth, tileHeight, prevX, prevY, greenToBlueCur, redToBlueCur, accumulatedBlueHisto); |
|
|
if (curDiff < bestDiff) |
|
|
if (curDiff < bestDiff) |
|
|
{ |
|
|
{ |
|
|
bestDiff = curDiff; |
|
|
bestDiff = curDiff; |
|
|
@ -917,9 +966,19 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
bestTx.RedToBlue = (byte)(redToBlueBest & 0xff); |
|
|
bestTx.RedToBlue = (byte)(redToBlueBest & 0xff); |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
private static double GetPredictionCostCrossColorRed(Span<uint> argb, int stride, int tileWidth, int tileHeight, Vp8LMultipliers prevX, Vp8LMultipliers prevY, int greenToRed, int[] accumulatedRedHisto) |
|
|
private static double GetPredictionCostCrossColorRed( |
|
|
|
|
|
Span<uint> argb, |
|
|
|
|
|
int stride, |
|
|
|
|
|
Span<int> scratch, |
|
|
|
|
|
int tileWidth, |
|
|
|
|
|
int tileHeight, |
|
|
|
|
|
Vp8LMultipliers prevX, |
|
|
|
|
|
Vp8LMultipliers prevY, |
|
|
|
|
|
int greenToRed, |
|
|
|
|
|
int[] accumulatedRedHisto) |
|
|
{ |
|
|
{ |
|
|
int[] histo = new int[256]; |
|
|
Span<int> histo = scratch.Slice(0, 256); |
|
|
|
|
|
histo.Clear(); |
|
|
|
|
|
|
|
|
CollectColorRedTransforms(argb, stride, tileWidth, tileHeight, greenToRed, histo); |
|
|
CollectColorRedTransforms(argb, stride, tileWidth, tileHeight, greenToRed, histo); |
|
|
double curDiff = PredictionCostCrossColor(accumulatedRedHisto, histo); |
|
|
double curDiff = PredictionCostCrossColor(accumulatedRedHisto, histo); |
|
|
@ -944,9 +1003,20 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
return curDiff; |
|
|
return curDiff; |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
private static double GetPredictionCostCrossColorBlue(Span<uint> argb, int stride, int tileWidth, int tileHeight, Vp8LMultipliers prevX, Vp8LMultipliers prevY, int greenToBlue, int redToBlue, int[] accumulatedBlueHisto) |
|
|
private static double GetPredictionCostCrossColorBlue( |
|
|
|
|
|
Span<uint> argb, |
|
|
|
|
|
int stride, |
|
|
|
|
|
Span<int> scratch, |
|
|
|
|
|
int tileWidth, |
|
|
|
|
|
int tileHeight, |
|
|
|
|
|
Vp8LMultipliers prevX, |
|
|
|
|
|
Vp8LMultipliers prevY, |
|
|
|
|
|
int greenToBlue, |
|
|
|
|
|
int redToBlue, |
|
|
|
|
|
int[] accumulatedBlueHisto) |
|
|
{ |
|
|
{ |
|
|
int[] histo = new int[256]; |
|
|
Span<int> histo = scratch.Slice(0, 256); |
|
|
|
|
|
histo.Clear(); |
|
|
|
|
|
|
|
|
CollectColorBlueTransforms(argb, stride, tileWidth, tileHeight, greenToBlue, redToBlue, histo); |
|
|
CollectColorBlueTransforms(argb, stride, tileWidth, tileHeight, greenToBlue, redToBlue, histo); |
|
|
double curDiff = PredictionCostCrossColor(accumulatedBlueHisto, histo); |
|
|
double curDiff = PredictionCostCrossColor(accumulatedBlueHisto, histo); |
|
|
@ -987,15 +1057,12 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
return curDiff; |
|
|
return curDiff; |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
private static void CollectColorRedTransforms(Span<uint> bgra, int stride, int tileWidth, int tileHeight, int greenToRed, int[] histo) |
|
|
private static void CollectColorRedTransforms(Span<uint> bgra, int stride, int tileWidth, int tileHeight, int greenToRed, Span<int> histo) |
|
|
{ |
|
|
{ |
|
|
#if SUPPORTS_RUNTIME_INTRINSICS
|
|
|
#if SUPPORTS_RUNTIME_INTRINSICS
|
|
|
if (Sse41.IsSupported) |
|
|
if (Sse41.IsSupported) |
|
|
{ |
|
|
{ |
|
|
var multsg = Vector128.Create(LosslessUtils.Cst5b(greenToRed)); |
|
|
var multsg = Vector128.Create(LosslessUtils.Cst5b(greenToRed)); |
|
|
var maskgreen = Vector128.Create(0x00ff00); |
|
|
|
|
|
var mask = Vector128.Create((short)0xff); |
|
|
|
|
|
|
|
|
|
|
|
const int span = 8; |
|
|
const int span = 8; |
|
|
Span<ushort> values = stackalloc ushort[span]; |
|
|
Span<ushort> values = stackalloc ushort[span]; |
|
|
for (int y = 0; y < tileHeight; y++) |
|
|
for (int y = 0; y < tileHeight; y++) |
|
|
@ -1011,15 +1078,15 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
uint* input1Idx = src + x + (span / 2); |
|
|
uint* input1Idx = src + x + (span / 2); |
|
|
Vector128<byte> input0 = Sse2.LoadVector128((ushort*)input0Idx).AsByte(); |
|
|
Vector128<byte> input0 = Sse2.LoadVector128((ushort*)input0Idx).AsByte(); |
|
|
Vector128<byte> input1 = Sse2.LoadVector128((ushort*)input1Idx).AsByte(); |
|
|
Vector128<byte> input1 = Sse2.LoadVector128((ushort*)input1Idx).AsByte(); |
|
|
Vector128<byte> g0 = Sse2.And(input0, maskgreen.AsByte()); // 0 0 | g 0
|
|
|
Vector128<byte> g0 = Sse2.And(input0, CollectColorRedTransformsGreenMask); // 0 0 | g 0
|
|
|
Vector128<byte> g1 = Sse2.And(input1, maskgreen.AsByte()); |
|
|
Vector128<byte> g1 = Sse2.And(input1, CollectColorRedTransformsGreenMask); |
|
|
Vector128<ushort> g = Sse41.PackUnsignedSaturate(g0.AsInt32(), g1.AsInt32()); // g 0
|
|
|
Vector128<ushort> g = Sse41.PackUnsignedSaturate(g0.AsInt32(), g1.AsInt32()); // g 0
|
|
|
Vector128<int> a0 = Sse2.ShiftRightLogical(input0.AsInt32(), 16); // 0 0 | x r
|
|
|
Vector128<int> a0 = Sse2.ShiftRightLogical(input0.AsInt32(), 16); // 0 0 | x r
|
|
|
Vector128<int> a1 = Sse2.ShiftRightLogical(input1.AsInt32(), 16); |
|
|
Vector128<int> a1 = Sse2.ShiftRightLogical(input1.AsInt32(), 16); |
|
|
Vector128<ushort> a = Sse41.PackUnsignedSaturate(a0, a1); // x r
|
|
|
Vector128<ushort> a = Sse41.PackUnsignedSaturate(a0, a1); // x r
|
|
|
Vector128<short> b = Sse2.MultiplyHigh(g.AsInt16(), multsg); // x dr
|
|
|
Vector128<short> b = Sse2.MultiplyHigh(g.AsInt16(), multsg); // x dr
|
|
|
Vector128<byte> c = Sse2.Subtract(a.AsByte(), b.AsByte()); // x r'
|
|
|
Vector128<byte> c = Sse2.Subtract(a.AsByte(), b.AsByte()); // x r'
|
|
|
Vector128<byte> d = Sse2.And(c, mask.AsByte()); // 0 r'
|
|
|
Vector128<byte> d = Sse2.And(c, CollectColorRedTransformsAndMask); // 0 r'
|
|
|
Sse2.Store(dst, d.AsUInt16()); |
|
|
Sse2.Store(dst, d.AsUInt16()); |
|
|
for (int i = 0; i < span; i++) |
|
|
for (int i = 0; i < span; i++) |
|
|
{ |
|
|
{ |
|
|
@ -1043,7 +1110,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
} |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
private static void CollectColorRedTransformsNoneVectorized(Span<uint> bgra, int stride, int tileWidth, int tileHeight, int greenToRed, int[] histo) |
|
|
private static void CollectColorRedTransformsNoneVectorized(Span<uint> bgra, int stride, int tileWidth, int tileHeight, int greenToRed, Span<int> histo) |
|
|
{ |
|
|
{ |
|
|
int pos = 0; |
|
|
int pos = 0; |
|
|
while (tileHeight-- > 0) |
|
|
while (tileHeight-- > 0) |
|
|
@ -1058,7 +1125,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
} |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
private static void CollectColorBlueTransforms(Span<uint> bgra, int stride, int tileWidth, int tileHeight, int greenToBlue, int redToBlue, int[] histo) |
|
|
private static void CollectColorBlueTransforms(Span<uint> bgra, int stride, int tileWidth, int tileHeight, int greenToBlue, int redToBlue, Span<int> histo) |
|
|
{ |
|
|
{ |
|
|
#if SUPPORTS_RUNTIME_INTRINSICS
|
|
|
#if SUPPORTS_RUNTIME_INTRINSICS
|
|
|
if (Sse41.IsSupported) |
|
|
if (Sse41.IsSupported) |
|
|
@ -1067,12 +1134,6 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
Span<ushort> values = stackalloc ushort[span]; |
|
|
Span<ushort> values = stackalloc ushort[span]; |
|
|
var multsr = Vector128.Create(LosslessUtils.Cst5b(redToBlue)); |
|
|
var multsr = Vector128.Create(LosslessUtils.Cst5b(redToBlue)); |
|
|
var multsg = Vector128.Create(LosslessUtils.Cst5b(greenToBlue)); |
|
|
var multsg = Vector128.Create(LosslessUtils.Cst5b(greenToBlue)); |
|
|
var maskgreen = Vector128.Create(0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255); |
|
|
|
|
|
var maskgreenblue = Vector128.Create(255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0, 255, 255, 0, 0); |
|
|
|
|
|
var maskblue = Vector128.Create(255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0, 255, 0); |
|
|
|
|
|
var shufflerLow = Vector128.Create(255, 2, 255, 6, 255, 10, 255, 14, 255, 255, 255, 255, 255, 255, 255, 255); |
|
|
|
|
|
var shufflerHigh = Vector128.Create(255, 255, 255, 255, 255, 255, 255, 255, 255, 2, 255, 6, 255, 10, 255, 14); |
|
|
|
|
|
|
|
|
|
|
|
for (int y = 0; y < tileHeight; y++) |
|
|
for (int y = 0; y < tileHeight; y++) |
|
|
{ |
|
|
{ |
|
|
Span<uint> srcSpan = bgra.Slice(y * stride); |
|
|
Span<uint> srcSpan = bgra.Slice(y * stride); |
|
|
@ -1086,18 +1147,18 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
uint* input1Idx = src + x + (span / 2); |
|
|
uint* input1Idx = src + x + (span / 2); |
|
|
Vector128<byte> input0 = Sse2.LoadVector128((ushort*)input0Idx).AsByte(); |
|
|
Vector128<byte> input0 = Sse2.LoadVector128((ushort*)input0Idx).AsByte(); |
|
|
Vector128<byte> input1 = Sse2.LoadVector128((ushort*)input1Idx).AsByte(); |
|
|
Vector128<byte> input1 = Sse2.LoadVector128((ushort*)input1Idx).AsByte(); |
|
|
Vector128<byte> r0 = Ssse3.Shuffle(input0, shufflerLow); |
|
|
Vector128<byte> r0 = Ssse3.Shuffle(input0, CollectColorBlueTransformsShuffleLowMask); |
|
|
Vector128<byte> r1 = Ssse3.Shuffle(input1, shufflerHigh); |
|
|
Vector128<byte> r1 = Ssse3.Shuffle(input1, CollectColorBlueTransformsShuffleHighMask); |
|
|
Vector128<byte> r = Sse2.Or(r0, r1); |
|
|
Vector128<byte> r = Sse2.Or(r0, r1); |
|
|
Vector128<byte> gb0 = Sse2.And(input0, maskgreenblue); |
|
|
Vector128<byte> gb0 = Sse2.And(input0, CollectColorBlueTransformsGreenBlueMask); |
|
|
Vector128<byte> gb1 = Sse2.And(input1, maskgreenblue); |
|
|
Vector128<byte> gb1 = Sse2.And(input1, CollectColorBlueTransformsGreenBlueMask); |
|
|
Vector128<ushort> gb = Sse41.PackUnsignedSaturate(gb0.AsInt32(), gb1.AsInt32()); |
|
|
Vector128<ushort> gb = Sse41.PackUnsignedSaturate(gb0.AsInt32(), gb1.AsInt32()); |
|
|
Vector128<byte> g = Sse2.And(gb.AsByte(), maskgreen); |
|
|
Vector128<byte> g = Sse2.And(gb.AsByte(), CollectColorBlueTransformsGreenMask); |
|
|
Vector128<short> a = Sse2.MultiplyHigh(r.AsInt16(), multsr); |
|
|
Vector128<short> a = Sse2.MultiplyHigh(r.AsInt16(), multsr); |
|
|
Vector128<short> b = Sse2.MultiplyHigh(g.AsInt16(), multsg); |
|
|
Vector128<short> b = Sse2.MultiplyHigh(g.AsInt16(), multsg); |
|
|
Vector128<byte> c = Sse2.Subtract(gb.AsByte(), b.AsByte()); |
|
|
Vector128<byte> c = Sse2.Subtract(gb.AsByte(), b.AsByte()); |
|
|
Vector128<byte> d = Sse2.Subtract(c, a.AsByte()); |
|
|
Vector128<byte> d = Sse2.Subtract(c, a.AsByte()); |
|
|
Vector128<byte> e = Sse2.And(d, maskblue); |
|
|
Vector128<byte> e = Sse2.And(d, CollectColorBlueTransformsBlueMask); |
|
|
Sse2.Store(dst, e.AsUInt16()); |
|
|
Sse2.Store(dst, e.AsUInt16()); |
|
|
for (int i = 0; i < span; i++) |
|
|
for (int i = 0; i < span; i++) |
|
|
{ |
|
|
{ |
|
|
@ -1121,7 +1182,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
} |
|
|
} |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
private static void CollectColorBlueTransformsNoneVectorized(Span<uint> bgra, int stride, int tileWidth, int tileHeight, int greenToBlue, int redToBlue, int[] histo) |
|
|
private static void CollectColorBlueTransformsNoneVectorized(Span<uint> bgra, int stride, int tileWidth, int tileHeight, int greenToBlue, int redToBlue, Span<int> histo) |
|
|
{ |
|
|
{ |
|
|
int pos = 0; |
|
|
int pos = 0; |
|
|
while (tileHeight-- > 0) |
|
|
while (tileHeight-- > 0) |
|
|
@ -1150,7 +1211,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
[MethodImpl(InliningOptions.ShortMethod)] |
|
|
[MethodImpl(InliningOptions.ShortMethod)] |
|
|
private static double PredictionCostCrossColor(int[] accumulated, int[] counts) |
|
|
private static double PredictionCostCrossColor(int[] accumulated, Span<int> counts) |
|
|
{ |
|
|
{ |
|
|
// Favor low entropy, locally and globally.
|
|
|
// Favor low entropy, locally and globally.
|
|
|
// Favor small absolute values for PredictionCostSpatial.
|
|
|
// Favor small absolute values for PredictionCostSpatial.
|
|
|
@ -1159,7 +1220,7 @@ namespace SixLabors.ImageSharp.Formats.Webp.Lossless |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
[MethodImpl(InliningOptions.ShortMethod)] |
|
|
[MethodImpl(InliningOptions.ShortMethod)] |
|
|
private static float PredictionCostSpatial(int[] counts, int weight0, double expVal) |
|
|
private static float PredictionCostSpatial(Span<int> counts, int weight0, double expVal) |
|
|
{ |
|
|
{ |
|
|
int significantSymbols = 256 >> 4; |
|
|
int significantSymbols = 256 >> 4; |
|
|
double expDecayFactor = 0.6; |
|
|
double expDecayFactor = 0.6; |
|
|
|
Brian Popow
5 years ago
GitHub