protected override void Compress4(ref byte[] block, int offset)
{
// build the table of lookups
var lookups = new[]
{
SingleColourLookupIns.Lookup54,
SingleColourLookupIns.Lookup64,
SingleColourLookupIns.Lookup54
};
// find the best end-points and index
ComputeEndPoints(lookups);
// build the block if we win
if (_mError >= _mBesterror)
{
return;
}
// remap the indices
var indices = new byte[16];
MColours.RemapIndices(new[] { _mIndex }, indices);
// save the block
ColourBlock.WriteColourBlock4(_mStart, _mEnd, indices, ref block, offset);
// save the error
_mBesterror = _mError;
}
protected override void Compress4(ref byte[] block, int offset)
{
// declare variables
var count = MColours.Count;
var two = new Vector4(2.0f);
var one = new Vector4(1.0f);
var onethirdOnethird2 = new Vector4(1.0f / 3.0f, 1.0f / 3.0f, 1.0f / 3.0f, 1.0f / 9.0f);
var twothirdsTwothirds2 = new Vector4(2.0f / 3.0f, 2.0f / 3.0f, 2.0f / 3.0f, 4.0f / 9.0f);
var twonineths = new Vector4(2.0f / 9.0f);
var zero = new Vector4(0.0f);
var half = new Vector4(0.5f);
var grid = new Vector4(31.0f, 63.0f, 31.0f, 0.0f);
var gridrcp = new Vector4(1.0f / 31.0f, 1.0f / 63.0f, 1.0f / 31.0f, 0.0f);
// prepare an ordering using the principle axis
ConstructOrdering(_mPrinciple, 0);
// check all possible clusters and iterate on the total order
var beststart = new Vector4(0.0f);
var bestend = new Vector4(0.0f);
var besterror = _mBesterror;
var bestindices = new byte[16];
var 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 (var iterationIndex = 0; ;)
{
// first cluster [0,i) is at the start
var part0 = new Vector4(0.0f);
for (var i = 0; i < count; ++i)
{
// second cluster [i,j) is one third along
var part1 = new Vector4(0.0f);
for (var j = i; ;)
{
// third cluster [j,k) is two thirds along
var part2 = j == 0 ? _mPointsWeights[0] : new Vector4(0.0f);
var kmin = j == 0 ? 1 : j;
for (var k = kmin; ;)
{
// last cluster [k,count) is at the end
var part3 = _mXsumWsum - part2 - part1 - part0;
// compute least squares terms directly
var alphaxSum = Helpers.MultiplyAdd(part2, onethirdOnethird2, Helpers.MultiplyAdd(part1, twothirdsTwothirds2, part0));
var alpha2Sum = alphaxSum.SplatW();
var betaxSum = Helpers.MultiplyAdd(part1, onethirdOnethird2, Helpers.MultiplyAdd(part2, twothirdsTwothirds2, part3));
var beta2Sum = betaxSum.SplatW();
var alphabetaSum = twonineths * (part1 + part2).SplatW();
// compute the least-squares optimal points
var factor = Helpers.Reciprocal(Helpers.NegativeMultiplySubtract(alphabetaSum, alphabetaSum, alpha2Sum * beta2Sum));
var a = Helpers.NegativeMultiplySubtract(betaxSum, alphabetaSum, alphaxSum * beta2Sum) * factor;
var b = Helpers.NegativeMultiplySubtract(alphaxSum, alphabetaSum, betaxSum * alpha2Sum) * factor;
// clamp to the grid
a = Vector4.Min(one, Vector4.Max(zero, a));
b = Vector4.Min(one, Vector4.Max(zero, b));
a = Helpers.Truncate(Helpers.MultiplyAdd(grid, a, half)) * gridrcp;
b = Helpers.Truncate(Helpers.MultiplyAdd(grid, b, half)) * gridrcp;
// compute the error (we skip the constant xxsum)
var e1 = Helpers.MultiplyAdd(a * a, alpha2Sum, b * b * beta2Sum);
var e2 = Helpers.NegativeMultiplySubtract(a, alphaxSum, a * b * alphabetaSum);
var e3 = Helpers.NegativeMultiplySubtract(b, betaxSum, e2);
var e4 = Helpers.MultiplyAdd(two, e3, e1);
// apply the metric to the error term
var e5 = e4 * _mMetric;
var error = e5.SplatX() + e5.SplatY() + e5.SplatZ();
// keep the solution if it wins
if (Helpers.CompareAnyLessThan(error, besterror))
{
beststart = a;
bestend = b;
besterror = error;
besti = i;
bestj = j;
bestk = k;
bestiteration = iterationIndex;
}
// advance
if (k == count)
{
break;
}
part2 += _mPointsWeights[k];
++k;
}
// advance
if (j == count)
{
break;
}
part1 += _mPointsWeights[j];
++j;
}
// advance
part0 += _mPointsWeights[i];
}
// stop if we didn't improve in this iteration
if (bestiteration != iterationIndex)
{
break;
}
// advance if possible
++iterationIndex;
if (iterationIndex == _mIterationCount)
{
break;
}
// stop if a new iteration is an ordering that has already been tried
var axis = (bestend - beststart).ToVector3();
if (!ConstructOrdering(axis, iterationIndex))
{
break;
}
}
// save the block if necessary
if (!Helpers.CompareAnyLessThan(besterror, _mBesterror))
{
return;
}
// remap the indices
var unordered = new byte[16];
for (var m = 0; m < besti; ++m)
{
unordered[_mOrder[16 * bestiteration + m]] = 0;
}
for (var m = besti; m < bestj; ++m)
{
unordered[_mOrder[16 * bestiteration + m]] = 2;
}
for (var m = bestj; m < bestk; ++m)
{
unordered[_mOrder[16 * bestiteration + m]] = 3;
}
for (var m = bestk; m < count; ++m)
{
unordered[_mOrder[16 * bestiteration + m]] = 1;
}
MColours.RemapIndices(unordered, bestindices);
// save the block
ColourBlock.WriteColourBlock4(beststart.ToVector3(), bestend.ToVector3(), bestindices, ref block, offset);
// save the error
_mBesterror = besterror;
}