Solving

A AMGX.Solver is created from a Resources, a Mode and a Config:

solver = AMGX.Solver(resources, AMGX.dDDI, config)

Solving is two phases. AMGX.setup! binds a matrix and builds the multigrid hierarchy — the expensive part — and AMGX.solve! then solves against a right-hand side:

x = AMGX.AMGXVector(resources, AMGX.dDDI)
AMGX.set_zero!(x, 3)

AMGX.setup!(solver, matrix)
AMGX.solve!(x, solver, rhs)

Vector(x)

The current contents of x are used as the initial guess; pass zero_inital_guess=true to ignore them.

Checking the result

Warning

A solve that does not converge is not an error — solve! returns normally and leaves a partial result in the solution vector. Always check the status.

AMGX.get_status(solver)

returns a AMGX.SolverStatus:

AMGX.SUCCESS — the convergence criterion was met.
AMGX.FAILED — the solver stopped on an internal error.
AMGX.DIVERGED — the solver reported divergence.
AMGX.NOT_CONVERGED — the criterion was not met, typically because max_iters was reached.

Iteration counts and residuals are available too:

AMGX.get_iterations_number(solver)
AMGX.get_iteration_residual(solver)     # final iteration
AMGX.get_iteration_residual(solver, 0)  # a specific iteration

Residuals for iterations other than the last require store_res_history=1 in the Config.

Re-solving with new coefficients

When only the matrix values changed, AMGX.resetup! reuses the existing hierarchy and is much cheaper than a second setup!:

AMGX.replace_coefficients!(matrix, new_values)
AMGX.resetup!(solver, matrix)
AMGX.solve!(x, solver, rhs)

resetup! requires the same matrix that setup! was given.