src/poisson.h

Multigrid Poisson–Helmholtz solvers

Multigrid Poisson–Helmholtz solvers

We want to solve Poisson–Helmholtz equations of the general form \displaystyle L(a) = \nabla\cdot (\alpha\nabla a) + \lambda a = b This can be done efficiently using a multigrid solver.

An important aspect of Poisson–Helmholtz equations is that the operator L() is linear. This property can be used to build better estimates of a solution by successive corrections to an initial guess. If we define an approximate solution \tilde{a} as \displaystyle \tilde{a} + da = a where a is the exact (unknown) solution, using the linearity of the operator we find that da verifies \displaystyle L(da) = b - L(\tilde{a}) where the right-hand-side is often called the residual of the approximate solution \tilde{a}.

Multigrid cycle

Here we implement the multigrid cycle proper. Given an initial guess a, a residual res, a correction field da and a relaxation function relax, we will provide an improved guess at the end of the cycle.

void mg_cycle (scalar * a, scalar * res, scalar * da,
	       void (* relax) (scalar * da, scalar * res, 
			       int depth, void * data),
	       void * data,
	       int nrelax, int minlevel, int maxlevel)
{

We first define the residual on all levels.

  restriction (res);

We then proceed from the coarsest grid (minlevel) down to the finest grid.

  minlevel = min (minlevel, maxlevel);
  for (int l = minlevel; l <= maxlevel; l++) {

On the coarsest grid, we take zero as initial guess.

    if (l == minlevel)
      foreach_level_or_leaf (l)
	for (scalar s in da)
	  foreach_blockf (s)
	    s[] = 0.;

On all other grids, we take as initial guess the approximate solution on the coarser grid bilinearly interpolated onto the current grid.

    else
      foreach_level (l)
	for (scalar s in da)
	  foreach_blockf (s)
	    s[] = bilinear (point, s);

We then apply homogeneous boundary conditions and do several iterations of the relaxation function to refine the initial guess.

    boundary_level (da, l);
    for (int i = 0; i < nrelax; i++) {
      relax (da, res, l, data);
      boundary_level (da, l);
    }
  }

And finally we apply the resulting correction to a.

  foreach() {
    scalar s, ds;
    for (s, ds in a, da)
      foreach_blockf (s)
	s[] += ds[];
  }
}

Multigrid solver

The multigrid solver itself uses successive calls to the multigrid cycle to refine an initial guess until a specified tolerance is reached.

The maximum number of iterations is controlled by NITERMAX and the tolerance by TOLERANCE with the default values below.

int NITERMAX = 100, NITERMIN = 1;
double TOLERANCE = 1e-3 [*];

Information about the convergence of the solver is returned in a structure.

typedef struct {
  int i;              // number of iterations
  double resb, resa;  // maximum residual before and after the iterations
  double sum;         // sum of r.h.s.
  int nrelax;         // number of relaxations
  int minlevel;       // minimum level of the multigrid hierarchy
} mgstats;

The user needs to provide a function which computes the residual field (and returns its maximum) as well as the relaxation function. The user-defined pointer data can be used to pass arguments to these functions. The optional number of relaxations is nrelax and res is an optional list of fields used to store the residuals. The minimum level of the hierarchy can be set (default is zero i.e. the root cell).

mgstats mg_solve (scalar * a, scalar * b,
		  double (* residual) (scalar * a, scalar * b, scalar * res,
				       void * data),
		  void (* relax) (scalar * da, scalar * res, int depth, 
				  void * data),
		  void * data = NULL,
		  int nrelax = 4,
		  scalar * res = NULL,
		  int minlevel = 0,
		  double tolerance = TOLERANCE)
{

We allocate a new correction and residual field for each of the scalars in a.

  scalar * da = list_clone (a), * pres = res;
  if (!res)
    res = list_clone (b);

The boundary conditions for the correction fields are the homogeneous equivalent of the boundary conditions applied to a.

  for (int b = 0; b < nboundary; b++)
    for (scalar s in da)
      s.boundary[b] = s.boundary_homogeneous[b];

We initialise the structure storing convergence statistics.

  mgstats s = {0};
  double sum = 0.;
  scalar rhs = b[0];
  foreach (reduction(+:sum))
    sum += rhs[];
  s.sum = sum;
  s.nrelax = nrelax > 0 ? nrelax : 4;

Here we compute the initial residual field and its maximum.

  double resb;
  resb = s.resb = s.resa = (* residual) (a, b, res, data);

We then iterate until convergence or until NITERMAX is reached. Note also that we force the solver to apply at least one cycle, even if the initial residual is lower than TOLERANCE.

  for (s.i = 0;
       s.i < NITERMAX && (s.i < NITERMIN || s.resa > tolerance);
       s.i++) {
    mg_cycle (a, res, da, relax, data,
	      s.nrelax,
	      minlevel,
	      grid->maxdepth);
    s.resa = (* residual) (a, b, res, data);

We tune the number of relaxations so that the residual is reduced by between 2 and 20 for each cycle. This is particularly useful for stiff systems which may require a larger number of relaxations on the finest grid.

#if 1
    if (s.resa > tolerance) {
      if (resb/s.resa < 1.2 && s.nrelax < 100)
	s.nrelax++;
      else if (resb/s.resa > 10 && s.nrelax > 2)
	s.nrelax--;
    }
#else
    if (s.resa == resb) /* convergence has stopped!! */
      break;
    if (s.resa > resb/1.1 && p.minlevel < grid->maxdepth)
      p.minlevel++;
#endif

    resb = s.resa;
  }
  s.minlevel = minlevel;

If we have not satisfied the tolerance, we warn the user.

  if (s.resa > tolerance) {
    scalar v = a[0]; // fixme: should not be necessary
    fprintf (ferr, 
	     "WARNING: convergence for %s not reached after %d iterations\n"
	     "  res: %g sum: %g nrelax: %d tolerance: %g\n", v.name,
	     s.i, s.resa, s.sum, s.nrelax, tolerance), fflush (ferr);
  }

We deallocate the residual and correction fields and free the lists.

  if (!pres)
    delete (res), free (res);
  delete (da), free (da);

  return s;
}

Application to the Poisson–Helmholtz equation

We now apply the generic multigrid solver to the Poisson–Helmholtz equation \displaystyle \nabla\cdot (\alpha\nabla a) + \lambda a = b We first setup the data structure required to pass the extra parameters \alpha and \lambda. We define \alpha as a face vector field because we need values at the face locations corresponding to the face gradients of field a.

alpha and lambda are declared as (const) to indicate that the function works also when alpha and lambda are constant vector (resp. scalar) fields. If tolerance is set, it supersedes the default TOLERANCE of the multigrid solver, nrelax controls the initial number of relaxations (default is one), minlevel controls the minimum level of the hierarchy (default is one) and res is an optional list of fields used to store the final residual (which can be useful to monitor convergence).

struct Poisson {
  scalar a, b;
  (const) face vector alpha;
  (const) scalar lambda;
  double tolerance;
  int nrelax, minlevel;
  scalar * res;
#if EMBED
  double (* embed_flux) (Point, scalar, vector, double *);
#endif
};

We can now write the relaxation function. We first recover the extra parameters from the data pointer.

static void relax (scalar * al, scalar * bl, int l, void * data)
{
  scalar a = al[0], b = bl[0];
  struct Poisson * p = (struct Poisson *) data;
  (const) face vector alpha = p->alpha;
  (const) scalar lambda = p->lambda;

We use either Jacobi (under)relaxation or we directly reuse values as soon as they are updated. For Jacobi, we need to allocate space for the new field c. Jacobi is useful mostly as it gives results which are independent of the order in which the cells are traversed. This is not the case for the simple traversal, which means for example that results will depend on whether a tree or a multigrid is used (because cells will be traversed in a different order). The same comment applies to OpenMP or MPI parallelism. In practice however Jacobi convergence tends to be slower than simple reuse.

#if JACOBI
  scalar c[];
#else
  scalar c = a;
#endif

We use the face values of \alpha to weight the gradients of the 5-points Laplacian operator. We get the relaxation function.

  foreach_level_or_leaf (l) {
    double n = - sq(Delta)*b[], d = - lambda[]*sq(Delta);
    foreach_dimension() {
      n += alpha.x[1]*a[1] + alpha.x[]*a[-1];
      d += alpha.x[1] + alpha.x[];
    }
#if EMBED
    if (p->embed_flux) {
      double c, e = p->embed_flux (point, a, alpha, &c);
      n -= c*sq(Delta);
      d += e*sq(Delta);
    }
    if (!d)
      c[] = 0., b[] = 0.;
    else
#endif // EMBED
      c[] = n/d;
  }

For weighted Jacobi we under-relax with a weight of 2/3.

#if JACOBI
  foreach_level_or_leaf (l)
    a[] = (a[] + 2.*c[])/3.;
#endif
  
#if TRASH
  scalar a1[];
  foreach_level_or_leaf (l)
    a1[] = a[];
  trash ({a});
  foreach_level_or_leaf (l)
    a[] = a1[];
#endif
}

The equivalent residual function is obtained in a similar way in the case of a Cartesian grid, however the case of the tree mesh requires more careful consideration…

static double residual (scalar * al, scalar * bl, scalar * resl, void * data)
{
  scalar a = al[0], b = bl[0], res = resl[0];
  struct Poisson * p = (struct Poisson *) data;
  (const) face vector alpha = p->alpha;
  (const) scalar lambda = p->lambda;
  double maxres = 0.;
#if TREE
  /* conservative coarse/fine discretisation (2nd order) */
  face vector g[];
  foreach_face()
    g.x[] = alpha.x[]*face_gradient_x (a, 0);
  foreach (reduction(max:maxres), nowarning) {
    res[] = b[] - lambda[]*a[];
    foreach_dimension()
      res[] -= (g.x[1] - g.x[])/Delta;
#if EMBED
    if (p->embed_flux) {
      double c, e = p->embed_flux (point, a, alpha, &c);
      res[] += c - e*a[];
    }
#endif // EMBED    
    if (fabs (res[]) > maxres)
      maxres = fabs (res[]);
  }
#else // !TREE
  /* "naive" discretisation (only 1st order on trees) */
  foreach (reduction(max:maxres), nowarning) {
    res[] = b[] - lambda[]*a[];
    foreach_dimension()
      res[] += (alpha.x[0]*face_gradient_x (a, 0) -
		alpha.x[1]*face_gradient_x (a, 1))/Delta;  
#if EMBED
    if (p->embed_flux) {
      double c, e = p->embed_flux (point, a, alpha, &c);
      res[] += c - e*a[];
    }
#endif // EMBED
    if (fabs (res[]) > maxres)
      maxres = fabs (res[]);
  }
#endif // !TREE
  return maxres;
}

User interface

Finally we provide a generic user interface for a Poisson–Helmholtz equation of the form \displaystyle \nabla\cdot (\alpha\nabla a) + \lambda a = b

mgstats poisson (scalar a, scalar b,
		 (const) face vector alpha = {{-1}},
		 (const) scalar lambda = {-1},
		 double tolerance = 0.,
		 int nrelax = 4,
		 int minlevel = 0,
		 scalar * res = NULL,
		 double (* flux) (Point, scalar, vector, double *) = NULL)
{

If \alpha or \lambda are not set, we replace them with constant unity vector (resp. zero scalar) fields. Note that the user is free to provide \alpha and \beta as constant fields.

  if (alpha.x.i < 0)
    alpha = unityf;
  if (lambda.i < 0) {
    const scalar zeroc[] = 0.; // fixme
    lambda = zeroc;
  }

We need \alpha and \lambda on all levels of the grid.

  restriction ({alpha,lambda});

If tolerance is set it supersedes the default of the multigrid solver.

  double defaultol = TOLERANCE;
  if (tolerance)
    TOLERANCE = tolerance;

  struct Poisson p = {a, b, alpha, lambda, tolerance, nrelax, minlevel, res };
#if EMBED
  if (!flux && a.boundary[embed] != symmetry)
    p.embed_flux = embed_flux;
  else
    p.embed_flux = flux;
#endif // EMBED
  mgstats s = mg_solve ({a}, {b}, residual, relax, &p,
			nrelax, res, max(1, minlevel));

We restore the default.

  if (tolerance)
    TOLERANCE = defaultol;

  return s;
}

Projection of a velocity field

The function below “projects” the velocity field u onto the space of divergence-free velocity fields i.e. \displaystyle \mathbf{u}_f^{n+1} \leftarrow \mathbf{u}_f - \Delta t\alpha\nabla p so that \displaystyle \nabla\cdot\mathbf{u}_f^{n+1} = 0 This gives the Poisson equation for the pressure \displaystyle \nabla\cdot(\alpha\nabla p) = \frac{\nabla\cdot\mathbf{u}_f}{\Delta t}

trace
mgstats project (face vector uf, scalar p,
		 (const) face vector alpha = unityf,
		 double dt = 1.,
		 int nrelax = 4)
{

We allocate a local scalar field and compute the divergence of \mathbf{u}_f. The divergence is scaled by dt so that the pressure has the correct dimension.

  scalar div[];
  foreach() {
    div[] = 0.;
    foreach_dimension()
      div[] += uf.x[1] - uf.x[];
    div[] /= dt*Delta;
  }

We solve the Poisson problem. The tolerance (set with TOLERANCE) is the maximum relative change in volume of a cell (due to the divergence of the flow) during one timestep i.e. the non-dimensional quantity \displaystyle |\nabla\cdot\mathbf{u}_f|\Delta t Given the scaling of the divergence above, this gives

  mgstats mgp = poisson (p, div, alpha,
			 tolerance = TOLERANCE/sq(dt), nrelax = nrelax);

And compute \mathbf{u}_f^{n+1} using \mathbf{u}_f and p.

  foreach_face()
    uf.x[] -= dt*alpha.x[]*face_gradient_x (p, 0);

  return mgp;
}