/// <summary>
/// Preconditioned variant
/// Similar to non-preconditioned version, this can suffer if one solution converges
/// much slower than others, as we can't skip matrix multiplies in that case.
/// </summary>
public bool SolvePreconditioned()
{
Iterations = 0;
if (B == null || MultiplyF == null || PreconditionMultiplyF == null)
{
throw new Exception("SparseSymmetricCGMultipleRHS.SolvePreconditioned(): Must set B and MultiplyF and PreconditionMultiplyF!");
}
int NRHS = B.Length;
if (NRHS == 0)
{
throw new Exception("SparseSymmetricCGMultipleRHS.SolvePreconditioned(): Need at least one RHS vector in B");
}
int n = B[0].Length;
R = BufferUtil.AllocNxM(NRHS, n);
P = BufferUtil.AllocNxM(NRHS, n);
AP = BufferUtil.AllocNxM(NRHS, n);
Z = BufferUtil.AllocNxM(NRHS, n);
if (X == null || UseXAsInitialGuess == false)
{
if (X == null)
{
X = BufferUtil.AllocNxM(NRHS, n);
}
for (int j = 0; j < NRHS; ++j)
{
Array.Clear(X[j], 0, n);
Array.Copy(B[j], R[j], n);
}
}
else
{
// hopefully is X is a decent initialization...
InitializeR(R);
}
// [RMS] for convergence test?
double[] norm = new double[NRHS];
for (int j = 0; j < NRHS; ++j)
{
norm[j] = BufferUtil.Dot(B[j], B[j]);
}
double[] root1 = new double[NRHS];
for (int j = 0; j < NRHS; ++j)
{
root1[j] = Math.Sqrt(norm[j]);
}
// r_0 = b - A*x_0
MultiplyF(X, R);
for (int j = 0; j < NRHS; ++j)
{
for (int i = 0; i < n; ++i)
{
R[j][i] = B[j][i] - R[j][i];
}
}
// z0 = M_inverse * r_0
PreconditionMultiplyF(R, Z);
// p0 = z0
for (int j = 0; j < NRHS; ++j)
{
Array.Copy(Z[j], P[j], n);
}
// compute initial R*Z
double[] RdotZ_k = new double[NRHS];
for (int j = 0; j < NRHS; ++j)
{
RdotZ_k[j] = BufferUtil.Dot(R[j], Z[j]);
}
double[] alpha_k = new double[NRHS];
double[] beta_k = new double[NRHS];
bool[] converged = new bool[NRHS];
var rhs = Interval1i.Range(NRHS);
int iter = 0;
while (iter++ < MaxIterations)
{
// convergence test
bool done = true;
for (int j = 0; j < NRHS; ++j)
{
if (converged[j] == false)
{
double root0 = Math.Sqrt(RdotZ_k[j]);
if (root0 <= ConvergeTolerance * root1[j])
{
converged[j] = true;
}
}
if (converged[j] == false)
{
done = false;
}
}
if (done)
{
break;
}
MultiplyF(P, AP);
gParallel.ForEach(rhs, (j) =>
{
if (converged[j] == false)
{
alpha_k[j] = RdotZ_k[j] / BufferUtil.Dot(P[j], AP[j]);
}
});
// x_k+1 = x_k + alpha_k * p_k
gParallel.ForEach(rhs, (j) =>
{
if (converged[j] == false)
{
BufferUtil.MultiplyAdd(X[j], alpha_k[j], P[j]);
}
});
// r_k+1 = r_k - alpha_k * A * p_k
gParallel.ForEach(rhs, (j) =>
{
if (converged[j] == false)
{
BufferUtil.MultiplyAdd(R[j], -alpha_k[j], AP[j]);
}
});
// z_k+1 = M_inverse * r_k+1
PreconditionMultiplyF(R, Z);
// beta_k = (z_k+1 * r_k+1) / (z_k * r_k)
gParallel.ForEach(rhs, (j) =>
{
if (converged[j] == false)
{
beta_k[j] = BufferUtil.Dot(Z[j], R[j]) / RdotZ_k[j];
}
});
// can do these in parallel but improvement is minimal
// p_k+1 = z_k+1 + beta_k * p_k
gParallel.ForEach(rhs, (j) =>
{
if (converged[j] == false)
{
for (int i = 0; i < n; ++i)
{
P[j][i] = Z[j][i] + beta_k[j] * P[j][i];
}
}
});
gParallel.ForEach(rhs, (j) =>
{
if (converged[j] == false)
{
RdotZ_k[j] = BufferUtil.Dot(R[j], Z[j]);
}
});
}
//System.Console.WriteLine("{0} iterations", iter);
Iterations = iter;
return(iter < MaxIterations);
}
/// <summary>
/// standard CG solve
/// </summary>
public bool Solve()
{
Iterations = 0;
if (B == null || MultiplyF == null)
{
throw new Exception("SparseSymmetricCGMultipleRHS.Solve(): Must set B and MultiplyF!");
}
int NRHS = B.Length;
if (NRHS == 0)
{
throw new Exception("SparseSymmetricCGMultipleRHS.Solve(): Need at least one RHS vector in B");
}
int size = B[0].Length;
// Based on the algorithm in "Matrix Computations" by Golum and Van Loan.
R = BufferUtil.AllocNxM(NRHS, size);
P = BufferUtil.AllocNxM(NRHS, size);
W = BufferUtil.AllocNxM(NRHS, size);
if (X == null || UseXAsInitialGuess == false)
{
if (X == null)
{
X = BufferUtil.AllocNxM(NRHS, size);
}
for (int j = 0; j < NRHS; ++j)
{
Array.Clear(X[j], 0, size);
Array.Copy(B[j], R[j], size);
}
}
else
{
// hopefully is X is a decent initialization...
InitializeR(R);
}
// [RMS] these were inside loop but they are constant!
double[] norm = new double[NRHS];
for (int j = 0; j < NRHS; ++j)
{
norm[j] = BufferUtil.Dot(B[j], B[j]);
}
double[] root1 = new double[NRHS];
for (int j = 0; j < NRHS; ++j)
{
root1[j] = Math.Sqrt(norm[j]);
}
// The first iteration.
double[] rho0 = new double[NRHS];
for (int j = 0; j < NRHS; ++j)
{
rho0[j] = BufferUtil.Dot(R[j], R[j]);
}
// [RMS] If we were initialized w/ constraints already satisfied,
// then we are done! (happens for example in mesh deformations)
bool[] converged = new bool[NRHS];
int nconverged = 0;
for (int j = 0; j < NRHS; ++j)
{
converged[j] = rho0[j] < (ConvergeTolerance * root1[j]);
if (converged[j])
{
nconverged++;
}
}
if (nconverged == NRHS)
{
return(true);
}
for (int j = 0; j < NRHS; ++j)
{
Array.Copy(R[j], P[j], size);
}
MultiplyF(P, W);
double[] alpha = new double[NRHS];
for (int j = 0; j < NRHS; ++j)
{
alpha[j] = rho0[j] / BufferUtil.Dot(P[j], W[j]);
}
for (int j = 0; j < NRHS; ++j)
{
BufferUtil.MultiplyAdd(X[j], alpha[j], P[j]);
}
for (int j = 0; j < NRHS; ++j)
{
BufferUtil.MultiplyAdd(R[j], -alpha[j], W[j]);
}
double[] rho1 = new double[NRHS];
for (int j = 0; j < NRHS; ++j)
{
rho1[j] = BufferUtil.Dot(R[j], R[j]);
}
double[] beta = new double[NRHS];
var rhs = Interval1i.Range(NRHS);
// The remaining iterations.
int iter;
for (iter = 1; iter < MaxIterations; ++iter)
{
bool done = true;
for (int j = 0; j < NRHS; ++j)
{
if (converged[j] == false)
{
double root0 = Math.Sqrt(rho1[j]);
if (root0 <= ConvergeTolerance * root1[j])
{
converged[j] = true;
}
}
if (converged[j] == false)
{
done = false;
}
}
if (done)
{
break;
}
for (int j = 0; j < NRHS; ++j)
{
beta[j] = rho1[j] / rho0[j];
}
UpdateP(P, beta, R, converged);
MultiplyF(P, W);
gParallel.ForEach(rhs, (j) =>
{
if (converged[j] == false)
{
alpha[j] = rho1[j] / BufferUtil.Dot(P[j], W[j]);
}
});
// can do all these in parallel, but improvement is minimal
gParallel.ForEach(rhs, (j) =>
{
if (converged[j] == false)
{
BufferUtil.MultiplyAdd(X[j], alpha[j], P[j]);
}
});
gParallel.ForEach(rhs, (j) =>
{
if (converged[j] == false)
{
rho0[j] = rho1[j];
rho1[j] = BufferUtil.MultiplyAdd_GetSqrSum(R[j], -alpha[j], W[j]);
}
});
}
//System.Console.WriteLine("{0} iterations", iter);
Iterations = iter;
return(iter < MaxIterations);
}