private static unsafe int GetEndPointSelectionErrorFast(ReadOnlySpan <uint> tile, int subsetCount, int partition, int w, int h, int maxError)
{
byte[] partitionTable = BC67Tables.PartitionTable[subsetCount - 1][partition];
Span <RgbaColor8> minColors = stackalloc RgbaColor8[subsetCount];
Span <RgbaColor8> maxColors = stackalloc RgbaColor8[subsetCount];
BC67Utils.GetMinMaxColors(partitionTable, tile, w, h, minColors, maxColors, subsetCount);
Span <uint> endPoints0 = stackalloc uint[subsetCount];
Span <uint> endPoints1 = stackalloc uint[subsetCount];
SelectEndPointsFast(partitionTable, tile, w, h, subsetCount, minColors, maxColors, endPoints0, endPoints1, uint.MaxValue);
Span <RgbaColor32> palette = stackalloc RgbaColor32[8];
int errorSum = 0;
for (int subset = 0; subset < subsetCount; subset++)
{
RgbaColor32 blockDir = maxColors[subset].GetColor32() - minColors[subset].GetColor32();
int sum = blockDir.R + blockDir.G + blockDir.B + blockDir.A;
if (sum != 0)
{
blockDir = (blockDir << 6) / new RgbaColor32(sum);
}
uint c0 = endPoints0[subset];
uint c1 = endPoints1[subset];
int pBit0 = GetPBit(c0, 6, 0);
int pBit1 = GetPBit(c1, 6, 0);
c0 = BC67Utils.Quantize(RgbaColor8.FromUInt32(c0), 6, 0, pBit0).ToUInt32();
c1 = BC67Utils.Quantize(RgbaColor8.FromUInt32(c1), 6, 0, pBit1).ToUInt32();
if (Sse41.IsSupported)
{
Vector128 <byte> c0Rep = Vector128.Create(c0).AsByte();
Vector128 <byte> c1Rep = Vector128.Create(c1).AsByte();
Vector128 <byte> c0c1 = Sse2.UnpackLow(c0Rep, c1Rep);
Vector128 <byte> rWeights;
Vector128 <byte> lWeights;
fixed(byte *pWeights = BC67Tables.Weights[1], pInvWeights = BC67Tables.InverseWeights[1])
{
rWeights = Sse2.LoadScalarVector128((ulong *)pWeights).AsByte();
lWeights = Sse2.LoadScalarVector128((ulong *)pInvWeights).AsByte();
}
Vector128 <byte> iWeights = Sse2.UnpackLow(rWeights, lWeights);
Vector128 <byte> iWeights01 = Sse2.UnpackLow(iWeights.AsInt16(), iWeights.AsInt16()).AsByte();
Vector128 <byte> iWeights23 = Sse2.UnpackHigh(iWeights.AsInt16(), iWeights.AsInt16()).AsByte();
Vector128 <byte> iWeights0 = Sse2.UnpackLow(iWeights01.AsInt16(), iWeights01.AsInt16()).AsByte();
Vector128 <byte> iWeights1 = Sse2.UnpackHigh(iWeights01.AsInt16(), iWeights01.AsInt16()).AsByte();
Vector128 <byte> iWeights2 = Sse2.UnpackLow(iWeights23.AsInt16(), iWeights23.AsInt16()).AsByte();
Vector128 <byte> iWeights3 = Sse2.UnpackHigh(iWeights23.AsInt16(), iWeights23.AsInt16()).AsByte();
private static Block EncodeFast(ReadOnlySpan <uint> tile, int w, int h)
{
(RgbaColor8 minColor, RgbaColor8 maxColor) = BC67Utils.GetMinMaxColors(tile, w, h);
bool alphaNotOne = minColor.A != 255 || maxColor.A != 255;
int variance = BC67Utils.SquaredDifference(minColor.GetColor32(), maxColor.GetColor32());
int selectedMode;
int indexMode = 0;
if (alphaNotOne)
{
bool constantAlpha = minColor.A == maxColor.A;
if (constantAlpha)
{
selectedMode = variance > MinColorVarianceForModeChange ? 7 : 6;
}
else
{
if (variance > MinColorVarianceForModeChange)
{
Span <uint> uniqueRGB = stackalloc uint[16];
Span <uint> uniqueAlpha = stackalloc uint[16];
int uniqueRGBCount = 0;
int uniqueAlphaCount = 0;
uint rgbMask = new RgbaColor8(255, 255, 255, 0).ToUInt32();
uint alphaMask = new RgbaColor8(0, 0, 0, 255).ToUInt32();
for (int i = 0; i < tile.Length; i++)
{
uint c = tile[i];
if (!uniqueRGB.Slice(0, uniqueRGBCount).Contains(c & rgbMask))
{
uniqueRGB[uniqueRGBCount++] = c & rgbMask;
}
if (!uniqueAlpha.Slice(0, uniqueAlphaCount).Contains(c & alphaMask))
{
uniqueAlpha[uniqueAlphaCount++] = c & alphaMask;
}
}
selectedMode = 4;
indexMode = uniqueRGBCount > uniqueAlphaCount ? 1 : 0;
}
else
{
selectedMode = 5;
}
}
}
else
{
if (variance > MinColorVarianceForModeChange)
{
selectedMode = 1;
}
else
{
selectedMode = 6;
}
}
int selectedPartition = 0;
if (selectedMode == 1 || selectedMode == 7)
{
int partitionSelectionLowestError = int.MaxValue;
for (int i = 0; i < _mostFrequentPartitions.Length; i++)
{
int p = _mostFrequentPartitions[i];
int error = GetEndPointSelectionErrorFast(tile, 2, p, w, h, partitionSelectionLowestError);
if (error < partitionSelectionLowestError)
{
partitionSelectionLowestError = error;
selectedPartition = p;
}
}
}
return(Encode(selectedMode, selectedPartition, 0, indexMode, fastMode: true, tile, w, h, out _));
}
private static Block Encode(
int mode,
int partition,
int rotation,
int indexMode,
bool fastMode,
ReadOnlySpan <uint> tile,
int w,
int h,
out int errorSum)
{
BC7ModeInfo modeInfo = BC67Tables.BC7ModeInfos[mode];
int subsetCount = modeInfo.SubsetCount;
int partitionBitCount = modeInfo.PartitionBitCount;
int rotationBitCount = modeInfo.RotationBitCount;
int indexModeBitCount = modeInfo.IndexModeBitCount;
int colorDepth = modeInfo.ColorDepth;
int alphaDepth = modeInfo.AlphaDepth;
int pBits = modeInfo.PBits;
int colorIndexBitCount = modeInfo.ColorIndexBitCount;
int alphaIndexBitCount = modeInfo.AlphaIndexBitCount;
bool separateAlphaIndices = alphaIndexBitCount != 0;
uint alphaMask;
if (separateAlphaIndices)
{
alphaMask = rotation switch
{
1 => new RgbaColor8(255, 0, 0, 0).ToUInt32(),
2 => new RgbaColor8(0, 255, 0, 0).ToUInt32(),
3 => new RgbaColor8(0, 0, 255, 0).ToUInt32(),
_ => new RgbaColor8(0, 0, 0, 255).ToUInt32()
};
}
else
{
alphaMask = new RgbaColor8(0, 0, 0, 0).ToUInt32();
}
if (indexMode != 0)
{
alphaMask = ~alphaMask;
}
//
// Select color palette.
//
Span <uint> endPoints0 = stackalloc uint[subsetCount];
Span <uint> endPoints1 = stackalloc uint[subsetCount];
SelectEndPoints(
tile,
w,
h,
endPoints0,
endPoints1,
subsetCount,
partition,
colorIndexBitCount,
colorDepth,
alphaDepth,
~alphaMask,
fastMode);
if (separateAlphaIndices)
{
SelectEndPoints(
tile,
w,
h,
endPoints0,
endPoints1,
subsetCount,
partition,
alphaIndexBitCount,
colorDepth,
alphaDepth,
alphaMask,
fastMode);
}
Span <int> pBitValues = stackalloc int[pBits];
for (int i = 0; i < pBits; i++)
{
int pBit;
if (pBits == subsetCount)
{
pBit = GetPBit(endPoints0[i], endPoints1[i], colorDepth, alphaDepth);
}
else
{
int subset = i >> 1;
uint color = (i & 1) == 0 ? endPoints0[subset] : endPoints1[subset];
pBit = GetPBit(color, colorDepth, alphaDepth);
}
pBitValues[i] = pBit;
}
int colorIndexCount = 1 << colorIndexBitCount;
int alphaIndexCount = 1 << alphaIndexBitCount;
Span <byte> colorIndices = stackalloc byte[16];
Span <byte> alphaIndices = stackalloc byte[16];
errorSum = BC67Utils.SelectIndices(
tile,
w,
h,
endPoints0,
endPoints1,
pBitValues,
colorIndices,
subsetCount,
partition,
colorIndexBitCount,
colorIndexCount,
colorDepth,
alphaDepth,
pBits,
alphaMask);
if (separateAlphaIndices)
{
errorSum += BC67Utils.SelectIndices(
tile,
w,
h,
endPoints0,
endPoints1,
pBitValues,
alphaIndices,
subsetCount,
partition,
alphaIndexBitCount,
alphaIndexCount,
colorDepth,
alphaDepth,
pBits,
~alphaMask);
}
Span <bool> colorSwapSubset = stackalloc bool[3];
for (int i = 0; i < 3; i++)
{
colorSwapSubset[i] = colorIndices[BC67Tables.FixUpIndices[subsetCount - 1][partition][i]] >= (colorIndexCount >> 1);
}
bool alphaSwapSubset = alphaIndices[0] >= (alphaIndexCount >> 1);
Block block = new Block();
int offset = 0;
block.Encode(1UL << mode, ref offset, mode + 1);
block.Encode((ulong)partition, ref offset, partitionBitCount);
block.Encode((ulong)rotation, ref offset, rotationBitCount);
block.Encode((ulong)indexMode, ref offset, indexModeBitCount);
for (int comp = 0; comp < 3; comp++)
{
int rotatedComp = comp;
if (((comp + 1) & 3) == rotation)
{
rotatedComp = 3;
}
for (int subset = 0; subset < subsetCount; subset++)
{
RgbaColor8 color0 = RgbaColor8.FromUInt32(endPoints0[subset]);
RgbaColor8 color1 = RgbaColor8.FromUInt32(endPoints1[subset]);
int pBit0 = -1, pBit1 = -1;
if (pBits == subsetCount)
{
pBit0 = pBit1 = pBitValues[subset];
}
else if (pBits != 0)
{
pBit0 = pBitValues[subset * 2];
pBit1 = pBitValues[subset * 2 + 1];
}
if (indexMode == 0 ? colorSwapSubset[subset] : alphaSwapSubset)
{
block.Encode(BC67Utils.QuantizeComponent(color1.GetComponent(rotatedComp), colorDepth, pBit1), ref offset, colorDepth);
block.Encode(BC67Utils.QuantizeComponent(color0.GetComponent(rotatedComp), colorDepth, pBit0), ref offset, colorDepth);
}
else
{
block.Encode(BC67Utils.QuantizeComponent(color0.GetComponent(rotatedComp), colorDepth, pBit0), ref offset, colorDepth);
block.Encode(BC67Utils.QuantizeComponent(color1.GetComponent(rotatedComp), colorDepth, pBit1), ref offset, colorDepth);
}
}
}
if (alphaDepth != 0)
{
int rotatedComp = (rotation - 1) & 3;
for (int subset = 0; subset < subsetCount; subset++)
{
RgbaColor8 color0 = RgbaColor8.FromUInt32(endPoints0[subset]);
RgbaColor8 color1 = RgbaColor8.FromUInt32(endPoints1[subset]);
int pBit0 = -1, pBit1 = -1;
if (pBits == subsetCount)
{
pBit0 = pBit1 = pBitValues[subset];
}
else if (pBits != 0)
{
pBit0 = pBitValues[subset * 2];
pBit1 = pBitValues[subset * 2 + 1];
}
if (separateAlphaIndices && indexMode == 0 ? alphaSwapSubset : colorSwapSubset[subset])
{
block.Encode(BC67Utils.QuantizeComponent(color1.GetComponent(rotatedComp), alphaDepth, pBit1), ref offset, alphaDepth);
block.Encode(BC67Utils.QuantizeComponent(color0.GetComponent(rotatedComp), alphaDepth, pBit0), ref offset, alphaDepth);
}
else
{
block.Encode(BC67Utils.QuantizeComponent(color0.GetComponent(rotatedComp), alphaDepth, pBit0), ref offset, alphaDepth);
block.Encode(BC67Utils.QuantizeComponent(color1.GetComponent(rotatedComp), alphaDepth, pBit1), ref offset, alphaDepth);
}
}
}
for (int i = 0; i < pBits; i++)
{
block.Encode((ulong)pBitValues[i], ref offset, 1);
}
byte[] fixUpTable = BC67Tables.FixUpIndices[subsetCount - 1][partition];
for (int i = 0; i < 16; i++)
{
int subset = BC67Tables.PartitionTable[subsetCount - 1][partition][i];
byte index = colorIndices[i];
if (colorSwapSubset[subset])
{
index = (byte)(index ^ (colorIndexCount - 1));
}
int finalIndexBitCount = i == fixUpTable[subset] ? colorIndexBitCount - 1 : colorIndexBitCount;
Debug.Assert(index < (1 << finalIndexBitCount));
block.Encode(index, ref offset, finalIndexBitCount);
}
if (separateAlphaIndices)
{
for (int i = 0; i < 16; i++)
{
byte index = alphaIndices[i];
if (alphaSwapSubset)
{
index = (byte)(index ^ (alphaIndexCount - 1));
}
int finalIndexBitCount = i == 0 ? alphaIndexBitCount - 1 : alphaIndexBitCount;
Debug.Assert(index < (1 << finalIndexBitCount));
block.Encode(index, ref offset, finalIndexBitCount);
}
}
return(block);
}
private static void DecodeBlock(Block block, Span <uint> output, int w, int h, int width)
{
int mode = BitOperations.TrailingZeroCount((byte)block.Low | 0x100);
if (mode == 8)
{
// Mode is invalid, the spec mandates that hardware fills the block with
// a transparent black color.
for (int ty = 0; ty < h; ty++)
{
int baseOffs = ty * width;
for (int tx = 0; tx < w; tx++)
{
int offs = baseOffs + tx;
output[offs] = 0;
}
}
return;
}
BC7ModeInfo modeInfo = BC67Tables.BC7ModeInfos[mode];
int offset = mode + 1;
int partition = (int)block.Decode(ref offset, modeInfo.PartitionBitCount);
int rotation = (int)block.Decode(ref offset, modeInfo.RotationBitCount);
int indexMode = (int)block.Decode(ref offset, modeInfo.IndexModeBitCount);
Debug.Assert(partition < 64);
Debug.Assert(rotation < 4);
Debug.Assert(indexMode < 2);
int endPointCount = modeInfo.SubsetCount * 2;
Span <RgbaColor32> endPoints = stackalloc RgbaColor32[endPointCount];
Span <byte> pValues = stackalloc byte[modeInfo.PBits];
endPoints.Fill(new RgbaColor32(0, 0, 0, 255));
for (int i = 0; i < endPointCount; i++)
{
endPoints[i].R = (int)block.Decode(ref offset, modeInfo.ColorDepth);
}
for (int i = 0; i < endPointCount; i++)
{
endPoints[i].G = (int)block.Decode(ref offset, modeInfo.ColorDepth);
}
for (int i = 0; i < endPointCount; i++)
{
endPoints[i].B = (int)block.Decode(ref offset, modeInfo.ColorDepth);
}
if (modeInfo.AlphaDepth != 0)
{
for (int i = 0; i < endPointCount; i++)
{
endPoints[i].A = (int)block.Decode(ref offset, modeInfo.AlphaDepth);
}
}
for (int i = 0; i < modeInfo.PBits; i++)
{
pValues[i] = (byte)block.Decode(ref offset, 1);
}
for (int i = 0; i < endPointCount; i++)
{
int pBit = -1;
if (modeInfo.PBits != 0)
{
int pIndex = (i * modeInfo.PBits) / endPointCount;
pBit = pValues[pIndex];
}
Unquantize(ref endPoints[i], modeInfo.ColorDepth, modeInfo.AlphaDepth, pBit);
}
byte[] partitionTable = BC67Tables.PartitionTable[modeInfo.SubsetCount - 1][partition];
byte[] fixUpTable = BC67Tables.FixUpIndices[modeInfo.SubsetCount - 1][partition];
Span <byte> colorIndices = stackalloc byte[16];
for (int i = 0; i < 16; i++)
{
byte subset = partitionTable[i];
int bitCount = i == fixUpTable[subset] ? modeInfo.ColorIndexBitCount - 1 : modeInfo.ColorIndexBitCount;
colorIndices[i] = (byte)block.Decode(ref offset, bitCount);
Debug.Assert(colorIndices[i] < 16);
}
Span <byte> alphaIndices = stackalloc byte[16];
if (modeInfo.AlphaIndexBitCount != 0)
{
for (int i = 0; i < 16; i++)
{
int bitCount = i != 0 ? modeInfo.AlphaIndexBitCount : modeInfo.AlphaIndexBitCount - 1;
alphaIndices[i] = (byte)block.Decode(ref offset, bitCount);
Debug.Assert(alphaIndices[i] < 16);
}
}
for (int ty = 0; ty < h; ty++)
{
int baseOffs = ty * width;
for (int tx = 0; tx < w; tx++)
{
int i = ty * 4 + tx;
RgbaColor32 color;
byte subset = partitionTable[i];
RgbaColor32 color1 = endPoints[subset * 2];
RgbaColor32 color2 = endPoints[subset * 2 + 1];
if (modeInfo.AlphaIndexBitCount != 0)
{
if (indexMode == 0)
{
color = BC67Utils.Interpolate(color1, color2, colorIndices[i], alphaIndices[i], modeInfo.ColorIndexBitCount, modeInfo.AlphaIndexBitCount);
}
else
{
color = BC67Utils.Interpolate(color1, color2, alphaIndices[i], colorIndices[i], modeInfo.AlphaIndexBitCount, modeInfo.ColorIndexBitCount);
}
}
else
{
color = BC67Utils.Interpolate(color1, color2, colorIndices[i], colorIndices[i], modeInfo.ColorIndexBitCount, modeInfo.ColorIndexBitCount);
}
if (rotation != 0)
{
int a = color.A;
switch (rotation)
{
case 1: color.A = color.R; color.R = a; break;
case 2: color.A = color.G; color.G = a; break;
case 3: color.A = color.B; color.B = a; break;
}
}
RgbaColor8 color8 = color.GetColor8();
output[baseOffs + tx] = color8.ToUInt32();
}
}
}