void PolarDecompositionGradient::Compute(const double * M, const double * Q, const double * S, const double * MDot, double * omega, double * QDot, double * SDot, const double * MDotDot, double * omegaDot, double * QDotDot)
{
// compute omega = G^{-1} (2 * skew(Q^T * MDot)), where G = (tr(S)I - S) * Q^T
// (see Barbic and Zhao, SIGGRAPH 2011)
// first, construct G, and invert it
// tempMatrix = tr(S)I - S
double tempMatrix[9];
for(int i=0; i<9; i++)
tempMatrix[i] = -S[i];
double trace = S[0] + S[4] + S[8];
tempMatrix[0] += trace;
tempMatrix[4] += trace;
tempMatrix[8] += trace;
double G[9]; // G = (tr(S)I - S) * Q^T
MATRIX_MULTIPLY3X3ABT(tempMatrix, Q, G);
Mat3d GM(G);
Mat3d GInvM = inv(GM);
double GInv[9];
GInvM.convertToArray(GInv);
// omega = GInv * (2 * skew(R^T * Mdot))
MATRIX_MULTIPLY3X3ATB(Q, MDot, tempMatrix);
double rhs[3];
SKEW_PART(tempMatrix, rhs);
VECTOR_SCALE3(rhs, 2.0);
MATRIX_VECTOR_MULTIPLY3X3(GInv, rhs, omega);
// compute QDot = tilde(omega) * Q
double omegaTilde[9];
SKEW_MATRIX(omega, omegaTilde);
//double QDot[9];
MATRIX_MULTIPLY3X3(omegaTilde, Q, QDot);
// compute SDot = Q^T * (MDot - QDot * S)
// tempMatrix = MDot - QDot * S
MATRIX_MULTIPLY3X3(QDot, S, tempMatrix);
for(int i=0; i<9; i++)
tempMatrix[i] = MDot[i] - tempMatrix[i];
// SDot = Q^T * tempMatrix
MATRIX_MULTIPLY3X3ATB(Q, tempMatrix, SDot);
if ((MDotDot != NULL) && (omegaDot != NULL))
{
// compute omegaDot = GInv * ( 2 skew(Q^T (ADotDot - omegaTilde * ADot)) - (tr(SDot) I - SDot) * Q^T * omega )
// (see Barbic and Zhao, SIGGRAPH 2011)
// tempMatrix = MDotDot - omegaTilde * MDot
MATRIX_MULTIPLY3X3(omegaTilde, MDot, tempMatrix);
for(int i=0; i<9; i++)
tempMatrix[i] = MDotDot[i] - tempMatrix[i];
double tempMatrix2[9];
// tempVector = 2 * skew(Q^T * tempMatrix)
MATRIX_MULTIPLY3X3ATB(Q, tempMatrix, tempMatrix2);
double tempVector[3];
SKEW_PART(tempMatrix2, tempVector);
VECTOR_SCALE3(tempVector, 2.0);
// tempMatrix = tr(SDot)I - SDot
for(int i=0; i<9; i++)
tempMatrix[i] = -SDot[i];
double trace = SDot[0] + SDot[4] + SDot[8];
tempMatrix[0] += trace;
tempMatrix[4] += trace;
tempMatrix[8] += trace;
// tempVector2 = (tempMatrix * Q^T) * omega
double tempVector2[3];
MATRIX_MULTIPLY3X3ABT(tempMatrix, Q, tempMatrix2);
MATRIX_VECTOR_MULTIPLY3X3(tempMatrix2, omega, tempVector2);
// tempVector -= tempVector2
VECTOR_SUBTRACTEQUAL3(tempVector, tempVector2);
// tempVector2 = GInv * tempVector
MATRIX_VECTOR_MULTIPLY3X3(GInv, tempVector, omegaDot);
if (QDotDot != NULL)
{
double tempMatrix[9];
SKEW_MATRIX(omegaDot, tempMatrix);
MATRIX_MULTIPLY3X3(omegaTilde, omegaTilde, tempMatrix2);
for(int i=0;i<9;i++)
tempMatrix[i] += tempMatrix2[i];
MATRIX_MULTIPLY3X3(tempMatrix, Q, QDotDot);
}
}
}
void CorotationalLinearFEM::ComputeForceAndStiffnessMatrixOfSubmesh(double * u, double * f, SparseMatrix * stiffnessMatrix, int warp, int elementLo, int elementHi)
{
// clear f to zero
if (f != NULL)
memset(f, 0, sizeof(double) * 3 * numVertices);
// clear stiffness matrix to zero
if (stiffnessMatrix != NULL)
stiffnessMatrix->ResetToZero();
for (int el=elementLo; el < elementHi; el++)
{
int vtxIndex[4];
for (int vtx=0; vtx<4; vtx++)
vtxIndex[vtx] = tetMesh->getVertexIndex(el, vtx);
double KElement[144]; // element stiffness matrix, to be computed below; row-major
if (warp > 0)
{
double P[16]; // the current world-coordinate positions (row-major)
/*
P = [ v0 v1 v2 v3 ]
[ 1 1 1 1 ]
*/
// rows 1,2,3
for(int i=0; i<3; i++)
for(int j=0; j<4; j++)
P[4 * i + j] = undeformedPositions[3 * vtxIndex[j] + i] + u[3 * vtxIndex[j] + i];
// row 4
for(int j=0; j<4; j++)
P[12 + j] = 1;
// F = P * Inverse(M)
double F[9]; // upper-left 3x3 block
for(int i=0; i<3; i++)
for(int j=0; j<3; j++)
{
F[3 * i + j] = 0;
for(int k=0; k<4; k++)
F[3 * i + j] += P[4 * i + k] * MInverse[el][4 * k + j];
}
double R[9]; // rotation (row-major)
double S[9]; // symmetric (row-major)
double tolerance = 1E-6;
int forceRotation = 1;
PolarDecomposition::Compute(F, R, S, tolerance, forceRotation);
// RK = R * K
// KElement = R * K * R^T
double RK[144]; // row-major
WarpMatrix(KElementUndeformed[el], R, RK, KElement);
// f = RK (RT x - x0)
double fElement[12];
for(int i=0; i<12; i++)
{
fElement[i] = 0;
for(int j=0; j<4; j++)
for(int l=0; l<3; l++)
fElement[i] += KElement[12 * i + 3 * j + l] * P[4 * l + j] - RK[12 * i + 3 * j + l] * undeformedPositions[3 * vtxIndex[j] + l];
}
// add fElement into the global f
if (f != NULL)
{
for(int j=0; j<4; j++)
for(int l=0; l<3; l++)
f[3 * vtxIndex[j] + l] += fElement[3 * j + l];
}
// compute exact stiffness matrix
if (warp == 2)
{
// compute G = (tr(S) I - S) R^T
double G[9];
double tr = S[0] + S[4] + S[8];
double temp[9];
for(int i=0; i<9; i++)
temp[i] = -S[i];
temp[0] += tr;
temp[4] += tr;
temp[8] += tr;
// G = temp * R^T
MATRIX_MULTIPLY3X3ABT(temp, R, G);
double invG[9]; // invG = G^{-1}
inverse3x3(G, invG);
double rhs[27]; // 3 x 9 matrix (column-major)
for(int i=0; i<3; i++)
for(int j=0; j<3; j++)
{
double temp[9];
for(int k=0; k<9; k++)
temp[k] = 0.0;
// copy i-th row of R into column j of temp
for(int k=0; k<3; k++)
temp[3 * k + j] = R[3 * i + k];
// extract the skew-symmetric part
SKEW_PART(temp, &rhs[3 * (3 * i + j)]);
}
// must undo division by 2 from inside the SKEW_PART macro
for(int i=0; i<27; i++)
rhs[i] *= 2.0;
// solve G * omega = rhs
double omega[27]; // column-major
for(int i=0; i<9; i++)
{
MATRIX_VECTOR_MULTIPLY3X3(invG, &rhs[3 * i], &omega[3 * i]);
}
double dRdF[81]; // each column is skew(omega) * R ; column-major
for(int i=0; i<9; i++)
{
double skew[9];
SKEW_MATRIX(&omega[3 * i], skew);
MATRIX_MULTIPLY3X3(skew, R, &dRdF[9 * i]);
}
double B[3][3][9];
// re-arrange dRdF into B, for easier dRdF * dFdx multiplication (to exploit sparsity of dFdx)
for(int i=0; i<3; i++)
for(int j=0; j<3; j++)
for(int k=0; k<3; k++)
for(int l=0; l<3; l++)
{
int row = 3 * i + k;
int column = 3 * j + l;
B[i][j][3 * k + l] = dRdF[9 * column + row];
}
// four pointers to a 3-vector
double * minv[4] = { &MInverse[el][0], &MInverse[el][4], &MInverse[el][8], &MInverse[el][12] }; // the four rows of MInverse (last column ignored)
double dRdx[108]; // derivative of the element rotation matrix with respect to the positions of the tet vertices; column-major
for(int k=0; k<4; k++)
for(int i=0; i<3; i++)
for(int j=0; j<3; j++)
{
double temp[3];
MATRIX_VECTOR_MULTIPLY3X3(B[i][j], minv[k], temp);
int row = 3 * i;
int column = 3 * k + j;
VECTOR_SET3(&dRdx[9 * column + row], temp);
}
// add contribution of dRdx to KElement
// term 1: \hat{dR/dxl} K (R^T x - m)
// compute K (R^T x - m)
double tempVec[12]; // R^T x - m
for(int vtx=0; vtx<4; vtx++)
{
double pos[3];
for(int i=0; i<3; i++)
pos[i] = P[4 * i + vtx];
MATRIX_VECTOR_MULTIPLY3X3T(R, pos, &tempVec[3*vtx]);
// subtract m
for(int i=0; i<3; i++)
tempVec[3*vtx+i] -= undeformedPositions[3 * vtxIndex[vtx] + i];
}
double a[12]; // a = K * tempVec
for (int i=0; i<12; i++)
{
a[i] = 0.0;
for (int j=0; j<12; j++)
a[i] += KElementUndeformed[el][12 * i + j] * tempVec[j];
}
// add [\hat{dR/dxl} K R^T x]_l, l=1 to 12
for(int column=0; column<12; column++)
{
double b[12]; // b = \hat{dR/dxl} * a
for(int j=0; j<4; j++)
{
MATRIX_VECTOR_MULTIPLY3X3(&dRdx[9 * column], &a[3*j], &b[3*j]);
}
// write b into KElement (add b to i-th column)
for(int row=0; row<12; row++)
KElement[12 * row + column] += b[row]; // KElement is row-major
}
// term 2: (R K \hat{dRdxl}^T)x
// re-write positions into a
for(int vtx=0; vtx<4; vtx++)
{
for(int i=0; i<3; i++)
a[3 * vtx + i] = P[4 * i + vtx];
}
// compute [\hat{dRdxl}^T x)]_l, l=1 to 12
for(int column=0; column<12; column++)
{
double b[12]; // b = \hat{dRdxl}^T * a
for(int j=0; j<4; j++)
{
MATRIX_VECTOR_MULTIPLY3X3T(&dRdx[9 * column], &a[3*j], &b[3*j]);
}
// add RK * b to column of KElement
int rowStart = 0;
for (int row=0; row<12; row++)
{
double contrib = 0.0;
for (int j=0; j<12; j++)
contrib += RK[rowStart + j] * b[j];
KElement[rowStart + column] += contrib;
rowStart += 12;
}
}
}
}
else
{
// no warp
memcpy(KElement, KElementUndeformed[el], sizeof(double) * 144);
// f = K u
double fElement[12];
for(int i=0; i<12; i++)
{
fElement[i] = 0;
for(int j=0; j<4; j++)
{
fElement[i] +=
KElement[12 * i + 3 * j + 0] * u[3 * vtxIndex[j] + 0] +
KElement[12 * i + 3 * j + 1] * u[3 * vtxIndex[j] + 1] +
KElement[12 * i + 3 * j + 2] * u[3 * vtxIndex[j] + 2];
}
}
// add fElement into the global f
if (f != NULL)
{
for(int j=0; j<4; j++)
{
f[3 * vtxIndex[j] + 0] += fElement[3 * j + 0];
f[3 * vtxIndex[j] + 1] += fElement[3 * j + 1];
f[3 * vtxIndex[j] + 2] += fElement[3 * j + 2];
}
}
}
if (stiffnessMatrix != NULL)
{
int * rowIndex = rowIndices[el];
int * columnIndex = columnIndices[el];
// add KElement to the global stiffness matrix
for (int i=0; i<4; i++)
for (int j=0; j<4; j++)
for(int k=0; k<3; k++)
for(int l=0; l<3; l++)
stiffnessMatrix->AddEntry(3 * rowIndex[i] + k, 3 * columnIndex[4 * i + j] + l, KElement[12 * (3 * i + k) + 3 * j + l]);
}
}
}
