protected override void Compress4(BlockWindow block)
{
// cache some values
var count = m_colours.Count;
var values = m_colours.Points;
// create a codebook
var codes = new Vec3[4];
codes[0] = m_start;
codes[1] = m_end;
codes[2] = (2.0f / 3.0f) * m_start + (1.0f / 3.0f) * m_end;
codes[3] = (1.0f / 3.0f) * m_start + (2.0f / 3.0f) * m_end;
// match each point to the closest code
var closest = new Byte[16];
float error = 0.0f;
for (int i = 0; i < count; ++i)
{
// find the closest code
float dist = float.MaxValue;
int idx = 0;
for (int j = 0; j < 4; ++j)
{
float d = (m_metric * (values[i] - codes[j])).LengthSquared();
if (d < dist)
{
dist = d;
idx = j;
}
}
// save the index
closest[i] = (Byte)idx;
// accumulate the error
error += dist;
}
// save this scheme if it wins
if (error < m_besterror)
{
// remap the indices
var indices = new Byte[16];
m_colours.RemapIndices(closest, indices);
// save the block
block.WriteColourBlock4(m_start, m_end, indices);
// save the error
m_besterror = error;
}
}
public void Compress(BlockWindow block)
{
if (block.IsDtx1)
{
if (m_colours.IsTransparent)
{
Compress3(block);
}
else
{
Compress4(block);
}
}
else
{
Compress4(block);
}
}
protected override void Compress4(BlockWindow block)
{
// find the best end-points and index
ComputeEndPoints(LOOKUPTABLES4);
// build the block if we win
if (m_error < m_besterror)
{
// remap the indices
var indices = new Byte[16];
m_colours.RemapIndices(new Byte[] { m_index }, indices);
// save the block
block.WriteColourBlock4(m_start, m_end, indices);
// save the error
m_besterror = m_error;
}
}
private static void DecompressImage(Bitmap dstImage, Byte[] blocks, CompressionMode mode)
{
// initialise the block input
var block = new BlockWindow(blocks, mode);
var targetRgba = new Byte[4 * 16];
// loop over blocks
for (int y = 0; y < dstImage.Height; y += 4)
{
for (int x = 0; x < dstImage.Width; x += 4)
{
// decompress the block
block.Decompress(targetRgba);
// write the decompressed pixels to the correct image locations
dstImage.SetBlock(x, y, targetRgba);
// advance
block.Offset += block.ByteLength;
}
}
}
protected abstract void Compress4(BlockWindow block);
protected override void Compress3(BlockWindow block)
{
// prepare an ordering using the principle axis
ConstructOrdering(m_principle);
// declare variables
int count = m_colours.Count;
Vec4 beststart = Vec4.Zero;
Vec4 bestend = Vec4.Zero;
var besterror = float.MaxValue;
var bestbesterror = float.MaxValue;
// check all possible clusters for this total order
var indices = new Byte[16];
var bestindices = new Byte[16];
// first cluster [0,i) is at the start
for (int m = 0; m < count; ++m)
{
indices[m] = 0;
m_alpha[m] = m_weights[m];
m_beta[m] = 0;
}
for (int i = count; i >= 0; --i)
{
// second cluster [i,j) is half along
for (int m = i; m < count; ++m)
{
indices[m] = 2;
m_alpha[m] = m_beta[m] = 0.5f * m_weights[m];
}
for (int j = count; j > i; --j)
{
// last cluster [j,k) is at the end
if (j < count)
{
indices[j] = 1;
m_alpha[j] = 0;
m_beta[j] = m_weights[j];
}
// solve a least squares problem to place the endpoints
var error = SolveLeastSquares(out Vec4 start, out Vec4 end);
// keep the solution if it wins
if (error < besterror)
{
beststart = start;
bestend = end;
indices.CopyTo(bestindices, 0);
besterror = error;
}
}
}
// save the block if necessary
if (besterror < bestbesterror)
{
// remap the indices
var unordered = new Byte[16];
for (int i = 0; i < count; ++i)
{
unordered[m_order[i]] = bestindices[i];
}
m_colours.RemapIndices(unordered, bestindices);
// save the block
block.WriteColourBlock3(beststart.GetVec3(), bestend.GetVec3(), bestindices);
// save the error
bestbesterror = besterror;
}
}
private static void CompressImage(Bitmap srcImage, Byte[] blocks, CompressionMode mode, CompressionOptions options)
{
// fix any bad flags
options = options.FixFlags();
int block_width = (srcImage.Width + 3) / 4;
int block_height = (srcImage.Height + 3) / 4;
// if the number of chunks to process is not very large, we better skip parallel processing
if (block_width * block_height < 16)
{
options &= ~CompressionOptions.UseParallelProcessing;
}
if ((options & CompressionOptions.UseParallelProcessing) != 0)
{
System.Threading.Tasks.Parallel.For
(
0,
block_height,
(y, state) =>
{
// initialise the block output
var block = new BlockWindow(blocks, mode);
block.Offset += block.ByteLength * y * block_width;
// build the 4x4 block of pixels
var sourceRgba = new Byte[16 * 4];
for (int x = 0; x < block_width; x++)
{
srcImage.CopyBlockTo(x * 4, y * 4, sourceRgba, out int mask);
// compress it into the output
block.CompressMasked(sourceRgba, mask, options);
// advance
block.Offset += block.ByteLength;
}
}
);
}
else
{
// initialise the block output
var block = new BlockWindow(blocks, mode);
// build the 4x4 block of pixels
var sourceRgba = new Byte[16 * 4];
// loop over blocks
for (int y = 0; y < block_height; ++y)
{
for (int x = 0; x < block_width; ++x)
{
srcImage.CopyBlockTo(x * 4, y * 4, sourceRgba, out int mask);
// compress it into the output
block.CompressMasked(sourceRgba, mask, options);
// advance
block.Offset += block.ByteLength;
}
}
}
}
protected override void Compress4(BlockWindow block)
{
// declare variables
var count = m_colours.Count;
// prepare an ordering using the principle axis
ConstructOrdering(m_principle, 0);
// check all possible clusters and iterate on the total order
Vec4 beststart = Vec4.Zero;
Vec4 bestend = Vec4.Zero;
float m_besterror = float.MaxValue;
float besterror = m_besterror;
var bestindices = new Byte[16];
int bestiteration = 0;
int besti = 0, bestj = 0, bestk = 0;
// loop over iterations (we avoid the case that all points in first or last cluster)
for (int iterationIndex = 0; ;)
{
// first cluster [0,i) is at the start
Vec4 part0 = Vec4.Zero;
for (int i = 0; i < count; ++i)
{
// second cluster [i,j) is one third along
Vec4 part1 = Vec4.Zero;
for (int j = i; ;)
{
// third cluster [j,k) is two thirds along
Vec4 part2 = (j == 0) ? m_points_weights[0] : Vec4.Zero;
int kmin = (j == 0) ? 1 : j;
for (int k = kmin; ;)
{
// last cluster [k,count) is at the end
Vec4 part3 = m_xsum_wsum - part2 - part1 - part0;
// compute least squares terms directly
Vec4 alphax_sum = ONETHIRD_ONETHIRD2.MultiplyAdd(part2, TWOTHIRDS_TWOTHIRDS2.MultiplyAdd(part1, part0));
Vec4 betax_sum = ONETHIRD_ONETHIRD2.MultiplyAdd(part1, TWOTHIRDS_TWOTHIRDS2.MultiplyAdd(part2, part3));
Vec4 alphabeta_sum = TWONINETHS * (part1 + part2).SplatW();
var error = ComputeLeastSquares(alphax_sum, betax_sum, alphabeta_sum, out Vec4 a, out Vec4 b);
// keep the solution if it wins
if (error < besterror)
{
beststart = a;
bestend = b;
besterror = error;
besti = i;
bestj = j;
bestk = k;
bestiteration = iterationIndex;
}
// advance
if (k == count)
{
break;
}
part2 += m_points_weights[k];
++k;
}
// advance
if (j == count)
{
break;
}
part1 += m_points_weights[j];
++j;
}
// advance
part0 += m_points_weights[i];
}
// stop if we didn't improve in this iteration
if (bestiteration != iterationIndex)
{
break;
}
// advance if possible
++iterationIndex;
if (iterationIndex == m_iterationCount)
{
break;
}
// stop if a new iteration is an ordering that has already been tried
Vec3 axis = (bestend - beststart).GetVec3();
if (!ConstructOrdering(axis, iterationIndex))
{
break;
}
}
// save the block if necessary
if (besterror < m_besterror)
{
// remap the indices
var orderIndex = 16 * bestiteration;
var unordered = new Byte[16];
for (int m = 0; m < besti; ++m)
{
unordered[m_order[orderIndex + m]] = 0;
}
for (int m = besti; m < bestj; ++m)
{
unordered[m_order[orderIndex + m]] = 2;
}
for (int m = bestj; m < bestk; ++m)
{
unordered[m_order[orderIndex + m]] = 3;
}
for (int m = bestk; m < count; ++m)
{
unordered[m_order[orderIndex + m]] = 1;
}
m_colours.RemapIndices(unordered, bestindices);
// save the block
block.WriteColourBlock4(beststart.GetVec3(), bestend.GetVec3(), bestindices);
// save the error
m_besterror = besterror;
}
}