gko::batch::log::BatchConvergence#

Batched analogue of gko::log::Convergence: records the iteration count at which each batch item converged (or stopped) and the residual norm at that point. Per-item arrays so you can inspect, for instance, that 30 of 32 items converged in fewer than 100 iterations and the remaining 2 hit the cap.

template<typename ValueType = default_precision>
class BatchConvergence #

Inherits from

Logs the final residuals and iteration counts for a batch solver.

The purpose of this logger is to give simple access to standard data generated by the solver once it has converged.

Note

The final logged residuals are the implicit residuals that have been computed within the solver process. Depending on the solver algorithm, this may be significantly different from the true residual (||b - Ax||).

Public Functions

inline const array<index_type> &get_num_iterations(
) const noexcept#
Returns:

The number of iterations for entire batch

inline const array<real_type> &get_residual_norm() const noexcept#
Returns:

The residual norms for the entire batch.

Public Static Functions

static inline std::unique_ptr<BatchConvergence> create(
const mask_type &enabled_events = gko::log::Logger::batch_solver_completed_mask,
)#

Creates a convergence logger. This dynamically allocates the memory, constructs the object and returns an std::unique_ptr to this object. TODO: See if the objects can be pre-allocated beforehand instead of being copied in the on_<> event

Parameters:

  • exec – the executor

  • enabled_events – the events enabled for this logger. By default all events.

Returns:

an std::unique_ptr to the the constructed object