Add a new iterative solver

Ginkgo iterative solvers follow a strict CRTP skeleton: a class template Solver<ValueType> that inherits from EnableLinOp<Solver>, plumbs a workspace through workspace_traits, and runs its iteration loop in apply_dense_impl. This page walks through the pattern using gko::solver::Cg (include/ginkgo/core/solver/cg.hpp, core/solver/cg.cpp) as the worked example.

What you commit to implement

File	What it holds
include/ginkgo/core/solver/my_solver.hpp	Class template, parameters_type, workspace_traits specialisation
core/solver/my_solver.cpp	apply_impl, apply_dense_impl, transpose/conj_transpose, workspace-trait method bodies, parse for JSON config
core/solver/my_solver_kernels.hpp	Per-step kernel declarations (e.g. CG has initialize + step_1 + step_2)
One backend implementation per kernel	reference/solver/, common/unified/solver/ or cuda/solver/, hip/solver/, omp/solver/, dpcpp/solver/
core/config/config_helper.hpp	Add an enum entry to LinOpFactoryType
core/config/registry.cpp	Add "solver::MySolver" to the configuration map
core/config/solver_config.cpp	Add the GKO_PARSE_VALUE_TYPE(MySolver, gko::solver::MySolver) specialisation
core/CMakeLists.txt	Add solver/my_solver.cpp to the source list

Kernel files follow the same rules as in Add a new kernel.

The class skeleton

template <typename ValueType = default_precision>
class MySolver
    : public EnableLinOp<MySolver<ValueType>>,
      public EnablePreconditionedIterativeSolver<ValueType, MySolver<ValueType>>,
      public Transposable {
    friend class EnableLinOp<MySolver>;
    friend class EnablePolymorphicObject<MySolver, LinOp>;
    GKO_ASSERT_SUPPORTED_VALUE_TYPE;

public:
    using value_type      = ValueType;
    using transposed_type = MySolver<ValueType>;     // self if symmetric

    std::unique_ptr<LinOp> transpose() const override;
    std::unique_ptr<LinOp> conj_transpose() const override;

    bool apply_uses_initial_guess() const override { return true; }

    class Factory;

    struct parameters_type
        : enable_preconditioned_iterative_solver_factory_parameters<
              parameters_type, Factory> {
        // Add extra solver-specific parameters here, e.g.:
        //     int GKO_FACTORY_PARAMETER_SCALAR(krylov_dim, 30);
    };

    GKO_ENABLE_LIN_OP_FACTORY(MySolver, parameters, Factory);
    GKO_ENABLE_BUILD_METHOD(Factory);

    static parameters_type parse(
        const config::pnode& config,
        const config::registry& context,
        const config::type_descriptor& td_for_child =
            config::make_type_descriptor<ValueType>());

protected:
    void apply_impl(const LinOp* b, LinOp* x) const override;

    template <typename VectorType>
    void apply_dense_impl(const VectorType* b, VectorType* x) const;

    void apply_impl(const LinOp* alpha, const LinOp* b,
                    const LinOp* beta, LinOp* x) const override;

    explicit MySolver(std::shared_ptr<const Executor> exec)
        : EnableLinOp<MySolver>(std::move(exec)) {}

    explicit MySolver(const Factory* factory,
                      std::shared_ptr<const LinOp> system_matrix)
        : EnableLinOp<MySolver>(factory->get_executor(),
                                gko::transpose(system_matrix->get_size())),
          EnablePreconditionedIterativeSolver<ValueType, MySolver>{
              std::move(system_matrix), factory->get_parameters()},
          parameters_{factory->get_parameters()} {}
};

Key pieces:

• EnableLinOp<MySolver> injects the public apply overloads — you only implement apply_impl.

• EnablePreconditionedIterativeSolver is a mixin that gives every preconditioned iterative solver with_criteria(...), with_preconditioner(...), with_generated_preconditioner(...), plus the corresponding getters. Subclassing it pulls the factory parameter trait in via enable_preconditioned_iterative_solver_factory_parameters.

• apply_uses_initial_guess() { return true; } tells the framework that x on entry to apply is the initial guess and must not be zeroed.

• GKO_ENABLE_LIN_OP_FACTORY and GKO_ENABLE_BUILD_METHOD together define the Factory class, parameters_, the build() method, and the with_* setters. They’re macro-heavy boilerplate; the parameters_type struct above is the one place you decide what new knobs the solver exposes.

workspace_traits — the typed scratch pool

Iterative solvers allocate the same temporary vectors and scalars every iteration. To avoid reallocating on every apply, the base SolverBaseLinOp holds a mutable workspace that caches them. The solver tells the workspace how many slots it needs, and what each slot is for, through a workspace_traits<MySolver<V>> specialisation:

template <typename ValueType>
struct workspace_traits<MySolver<ValueType>> {
    using Solver = MySolver<ValueType>;
    static int num_vectors(const Solver&);
    static int num_arrays(const Solver&);
    static std::vector<std::string> op_names(const Solver&);
    static std::vector<std::string> array_names(const Solver&);
    static std::vector<int> scalars(const Solver&);
    static std::vector<int> vectors(const Solver&);

    // Slot IDs for operators (Dense* vectors / scalars):
    constexpr static int r        = 0;  // residual
    constexpr static int z        = 1;  // preconditioned residual
    constexpr static int p        = 2;
    constexpr static int q        = 3;  // A*p
    constexpr static int beta     = 4;
    constexpr static int prev_rho = 5;
    constexpr static int rho      = 6;
    constexpr static int one      = 7;  // constant 1.0
    constexpr static int minus_one = 8; // constant -1.0

    // Slot IDs for typed arrays:
    constexpr static int stop = 0;  // stopping_status array
    constexpr static int tmp  = 1;  // reduction scratch (char)
};

The trait method bodies in .cpp just return constants and string lists; e.g. CG returns 9 vectors and 2 arrays. The names appear in log output, so make them descriptive.

scalars() returns the IDs of the slots that are 1×nrhs (independent of problem size), vectors() returns the IDs of the slots that are n×nrhs (allocated per problem). The base class uses this to decide when to reallocate and when to reuse.

apply_impl and the precision dispatch

The two apply_impl overrides dispatch precision and then delegate to the templated apply_dense_impl. Use the distributed-aware variant — it accepts both dist::Vector and matrix::Dense, so the same code path works in serial and MPI builds:

template <typename ValueType>
void MySolver<ValueType>::apply_impl(const LinOp* b, LinOp* x) const
{
    if (!this->get_system_matrix()) {
        return;
    }
    experimental::precision_dispatch_real_complex_distributed<ValueType>(
        [this](auto dense_b, auto dense_x) {
            this->apply_dense_impl(dense_b, dense_x);
        },
        b, x);
}

template <typename ValueType>
void MySolver<ValueType>::apply_impl(const LinOp* alpha, const LinOp* b,
                                     const LinOp* beta, LinOp* x) const
{
    experimental::precision_dispatch_real_complex_distributed<ValueType>(
        [this](auto dense_alpha, auto dense_b, auto dense_beta, auto dense_x) {
            auto x_clone = dense_x->clone();
            this->apply_dense_impl(dense_b, x_clone.get());
            dense_x->scale(dense_beta);
            dense_x->add_scaled(dense_alpha, x_clone);
        },
        alpha, b, beta, x);
}

If your solver is not distributed-ready, use precision_dispatch_real_complex<ValueType> instead — it accepts only matrix::Dense.

apply_dense_impl — the iteration loop

This is the actual algorithm. The boilerplate macros from core/solver/solver_boilerplate.hpp allocate the workspace slots into named local variables of type Dense<V>*:

template <typename ValueType>
template <typename VectorType>
void MySolver<ValueType>::apply_dense_impl(const VectorType* dense_b,
                                           VectorType* dense_x) const
{
    using std::swap;
    constexpr uint8 RelativeStoppingId{1};

    auto exec = this->get_executor();
    this->setup_workspace();

    GKO_SOLVER_VECTOR(r, dense_b);
    GKO_SOLVER_VECTOR(z, dense_b);
    GKO_SOLVER_VECTOR(p, dense_b);
    GKO_SOLVER_VECTOR(q, dense_b);

    GKO_SOLVER_SCALAR(beta, dense_b);
    GKO_SOLVER_SCALAR(prev_rho, dense_b);
    GKO_SOLVER_SCALAR(rho, dense_b);

    GKO_SOLVER_ONE_MINUS_ONE();
    GKO_SOLVER_STOP_REDUCTION_ARRAYS();
    bool one_changed{};

    // r = b - A*x;  rho = 0;  prev_rho = 1;  z, p, q = 0
    exec->run(my_solver::make_initialize(
        gko::detail::get_local(dense_b)->get_const_device_view(),
        gko::detail::get_local(r)->get_device_view(),
        /* ... */ stop_status));
    this->get_system_matrix()->apply(neg_one_op, dense_x, one_op, r);

    auto stop_criterion = this->get_stop_criterion_factory()->generate(
        this->get_system_matrix(),
        std::shared_ptr<const LinOp>(dense_b, [](const LinOp*) {}),
        dense_x, r);

    int iter = -1;
    while (true) {
        this->get_preconditioner()->apply(r, z);
        r->compute_conj_dot(z, rho, reduction_tmp);

        ++iter;
        bool all_stopped =
            stop_criterion->update()
                .num_iterations(iter)
                .residual(r)
                .implicit_sq_residual_norm(rho)
                .solution(dense_x)
                .check(RelativeStoppingId, true,
                       &stop_status, &one_changed);
        this->template log<log::Logger::iteration_complete>(
            this, dense_b, dense_x, iter, r, nullptr, rho,
            &stop_status, all_stopped);
        if (all_stopped) {
            break;
        }

        // ... rest of the iteration steps as kernels ...
        swap(prev_rho, rho);
    }
}

Boilerplate macros (from core/solver/solver_boilerplate.hpp):

Macro	Effect
GKO_SOLVER_VECTOR(name, tmpl)	Allocates a Dense* matching tmpl’s shape into a local name.
GKO_SOLVER_SCALAR(name, tmpl)	Allocates a 1×nrhs Dense* (ValueType) into name.
GKO_SOLVER_ONE_MINUS_ONE()	Creates the constants one_op = 1 and neg_one_op = −1.
GKO_SOLVER_STOP_REDUCTION_ARRAYS()	Allocates the stop_status and reduction_tmp array<...>s.

gko::detail::get_local(v) extracts the rank-local matrix::Dense<V>* from a vector pointer — works for both experimental::distributed::Vector and ordinary matrix::Dense, so the same iteration body covers both cases.

transpose, conj_transpose

When the solver is symmetric (CG, MINRES, Chebyshev), transposed_type = MySolver<ValueType> and transpose() rebuilds the solver around the transposed system matrix and transposed preconditioner. Copy the body from Cg::transpose / Cg::conj_transpose in core/solver/cg.cpp and adjust the operator factory. If your solver is not transposable, do not include public Transposable in the base list.

Parse and register for JSON config

To make the solver constructible from a pnode / JSON tree, three edits are needed:

1. core/config/config_helper.hpp — append MySolver to the LinOpFactoryType enum.

2. core/config/registry.cpp — add an entry to generate_config_map():

   {"solver::MySolver", parse<LinOpFactoryType::MySolver>},
   

3. core/config/solver_config.cpp — add the parse specialisation:

   GKO_PARSE_VALUE_TYPE(MySolver, gko::solver::MySolver);
   

   This expands into the explicit template <> parse<LinOpFactoryType::MySolver> specialisation. For a value-and-index-templated solver, use GKO_PARSE_VALUE_AND_INDEX_TYPE. The Direct solver and the triangular solvers are the existing precedents.

The corresponding MySolver::parse(...) body inside core/solver/my_solver.cpp is one-liner — let common_solver_parse do the work:

template <typename ValueType>
typename MySolver<ValueType>::parameters_type MySolver<ValueType>::parse(
    const config::pnode& config, const config::registry& context,
    const config::type_descriptor& td_for_child)
{
    auto params = MySolver<ValueType>::build();
    config::config_check_decorator config_check(config);
    config::common_solver_parse(params, config_check, context, td_for_child);
    return params;
}

If your solver has parameters beyond the preconditioned-iterative-solver defaults, parse those out of config_check between the common_solver_parse call and the return.

Instantiation

End the .cpp with the value-type-explicit instantiation block:

#define GKO_DECLARE_MY_SOLVER(ValueType)        class MySolver<ValueType>
#define GKO_DECLARE_MY_SOLVER_TRAITS(ValueType) \
    struct workspace_traits<MySolver<ValueType>>
GKO_INSTANTIATE_FOR_EACH_VALUE_TYPE(GKO_DECLARE_MY_SOLVER);
GKO_INSTANTIATE_FOR_EACH_VALUE_TYPE(GKO_DECLARE_MY_SOLVER_TRAITS);

Both lines are required — the second is what makes workspace_traits<MySolver<double>>::num_vectors(...) actually link.

Reference-parity test

The full test plan is covered in Write tests. For a new solver you write three files:

• core/test/solver/my_solver.cpp — pure-C++ tests on the factory: parameters_type defaults, with_* setters, build() accepting generated preconditioners, transpose() returning a valid object. These don’t need a system matrix or iteration. Registered with ginkgo_create_test(my_solver) in core/test/solver/CMakeLists.txt.

• reference/test/solver/my_solver_kernels.cpp — Reference-only convergence and correctness tests on a small SPD or general system (depending on the solver class). Registered with ginkgo_create_test(my_solver_kernels) in reference/test/solver/CMakeLists.txt.

• test/solver/my_solver_kernels.cpp — the cross-backend parity test. Same source file compiled once per enabled backend (omp, cuda, hip, dpcpp) by ginkgo_create_common_test, using CommonTestFixture (from test/utils/common_fixture.hpp) which provides both ref (Reference) and exec (the target backend). Each test sets up the same system on both, runs the solver, and asserts the solutions agree with GKO_ASSERT_MTX_NEAR(...). Registered with ginkgo_create_common_test(my_solver_kernels) in test/solver/CMakeLists.txt.

cg_kernels.cpp in each of those three trees is the existing template to copy from.

See also

• Add a new kernel — for the iteration-step kernels.

• Add a new preconditioner — the simpler no-workspace variant of this pattern.

• Configure via JSON — what the registry edits above unlock for the end user.