Tuning algorithm for P3M. More...

#include <TuningAlgorithm.hpp>

Inheritance diagram for TuningAlgorithm:

Collaboration diagram for TuningAlgorithm:

Classes
struct	Parameters

Public Member Functions
	TuningAlgorithm (System::System &system, double prefactor, int timings)

virtual	~TuningAlgorithm ()=default

virtual P3MParameters &	get_params ()=0
	Get the P3M parameters.

virtual void	on_solver_change () const =0
	Re-initialize the currently active solver.

virtual TuningAlgorithm::Parameters	get_time ()=0
	Tuning loop entry point.

virtual void	setup_logger (bool verbose)=0
	Configure the logger.

virtual void	determine_mesh_limits ()=0
	Determine a sensible range for the mesh.

void	determine_r_cut_limits ()
	Determine a sensible range for the real-space cutoff.

void	determine_cao_limits (int initial_cao)
	Determine a sensible range for the charge assignment order.

virtual std::tuple< double, double, double, double >	calculate_accuracy (Utils::Vector3i const &mesh, int cao, double r_cut_iL) const =0
	Get the minimal error for this combination of parameters.

virtual std::optional< std::string >	layer_correction_veto_r_cut (double r_cut) const =0
	Veto real-space cutoffs larger than the layer correction gap.

virtual std::optional< std::string >	fft_decomposition_veto (Utils::Vector3i const &) const
	Veto FFT decomposition in non-cubic boxes.

void	commit (Utils::Vector3i const &mesh, int cao, double r_cut_iL, double alpha_L)
	Write tuned parameters to the P3M parameter struct.

void	tune ()

Protected Member Functions
auto	get_n_trials ()

void	increment_n_trials ()

void	reset_n_trials ()

double	get_m_time (Utils::Vector3i const &mesh, int &tuned_cao, double &tuned_r_cut_iL, double &tuned_alpha_L, double &tuned_accuracy)
	Get the optimal alpha and the corresponding computation time for a fixed `mesh`.

double	get_mc_time (Utils::Vector3i const &mesh, int cao, double &tuned_r_cut_iL, double &tuned_alpha_L, double &tuned_accuracy)
	Get the optimal alpha and the corresponding computation time for a fixed `mesh` and `cao`.

Protected Attributes
System::System &	m_system

double	m_prefactor

std::unique_ptr< TuningLogger >	m_logger = nullptr

int	cao_min = -1

int	cao_max = -1

int	cao_best = -1

double	m_r_cut_iL_min = -1.

double	m_r_cut_iL_max = -1.

Static Protected Attributes
static auto constexpr	time_granularity = 2.
	Granularity of the time measurement (milliseconds).

static auto constexpr	max_n_consecutive_trials = 20
	Maximal number of consecutive trials that don't improve runtime.

static auto constexpr	time_sentinel = std::numeric_limits<double>::max()
	Value for invalid time measurements.

Detailed Description

Tuning algorithm for P3M.

The algorithm basically determines the mesh, cao and then the real-space cutoff, in this order.

For each mesh, the optimal cao for the previous mesh is re-used as an initial guess, and the algorithm checks whether increasing or decreasing it leads to a better solution. This is efficient, since the optimal cao only changes little with the meshes in general.

The real-space cutoff for a given mesh and cao is determined via a bisection on the error estimate, which determines where the error estimate equals the required accuracy. Therefore the smallest possible, i.e. fastest real-space cutoff is determined.

Both the search over mesh and cao stop to search in a specific direction once the computation time is significantly higher than the currently known optimum.

Definition at line 64 of file TuningAlgorithm.hpp.

Constructor & Destructor Documentation

◆ TuningAlgorithm()

TuningAlgorithm::TuningAlgorithm	(	System::System &	system,
		double	prefactor,
		int	timings
	)

inline

Definition at line 94 of file TuningAlgorithm.hpp.

◆ ~TuningAlgorithm()

virtual TuningAlgorithm::~TuningAlgorithm ( )

virtualdefault

Member Function Documentation

◆ calculate_accuracy()

virtual std::tuple< double, double, double, double > TuningAlgorithm::calculate_accuracy	(	Utils::Vector3i const &	mesh,
		int	cao,
		double	r_cut_iL
	)		const

pure virtual

Get the minimal error for this combination of parameters.

The real-space error is tuned such that it contributes half of the total error, and then the k-space error is calculated. If an optimal alpha is not found, the value 0.1 is used as fallback.

Parameters

[in]	mesh	number of mesh points per coordinate direction (>0), in real space.
[in]	cao	charge assignment order ([0,7]).
[in]	r_cut_iL	cutoff radius for real space electrostatics (>0), rescaled to `r_cut_iL` = `r_cut` * `box_l_i`.

Returns: Error magnitude, real-space error, k-space error,

Ewald splitting parameter (0.

Implemented in CoulombTuningAlgorithm< FloatType, Architecture >, and DipolarTuningAlgorithm< FloatType, Architecture >.

Referenced by get_mc_time().

◆ commit()

void TuningAlgorithm::commit	(	Utils::Vector3i const &	mesh,
		int	cao,
		double	r_cut_iL,
		double	alpha_L
	)

Write tuned parameters to the P3M parameter struct.

Definition at line 90 of file TuningAlgorithm.cpp.

References System::System::box_geo, get_params(), and m_system.

Referenced by get_mc_time(), and tune().

◆ determine_cao_limits()

void TuningAlgorithm::determine_cao_limits ( int initial_cao )

Determine a sensible range for the charge assignment order.

Definition at line 77 of file TuningAlgorithm.cpp.

References P3MParameters::cao, cao_best, cao_max, cao_min, get_params(), and m_logger.

Referenced by CoulombP3MImpl< FloatType, Architecture >::tune(), and DipolarP3MImpl< FloatType, Architecture >::tune().

◆ determine_mesh_limits()

virtual void TuningAlgorithm::determine_mesh_limits ( )

pure virtual

Determine a sensible range for the mesh.

Implemented in CoulombTuningAlgorithm< FloatType, Architecture >, and DipolarTuningAlgorithm< FloatType, Architecture >.

◆ determine_r_cut_limits()

void TuningAlgorithm::determine_r_cut_limits ( )

Determine a sensible range for the real-space cutoff.

Definition at line 59 of file TuningAlgorithm.cpp.

References System::System::box_geo, System::System::cell_structure, get_params(), System::System::local_geo, m_logger, m_r_cut_iL_max, m_r_cut_iL_min, m_system, and P3MParameters::r_cut_iL.

Referenced by CoulombP3MImpl< FloatType, Architecture >::tune(), and DipolarP3MImpl< FloatType, Architecture >::tune().

◆ fft_decomposition_veto()

virtual std::optional< std::string > TuningAlgorithm::fft_decomposition_veto ( Utils::Vector3i const & ) const

inlinevirtual

Veto FFT decomposition in non-cubic boxes.

Reimplemented in CoulombTuningAlgorithm< FloatType, Architecture >.

Definition at line 152 of file TuningAlgorithm.hpp.

Referenced by get_mc_time().

◆ get_m_time()

double TuningAlgorithm::get_m_time	(	Utils::Vector3i const &	mesh,
		int &	tuned_cao,
		double &	tuned_r_cut_iL,
		double &	tuned_alpha_L,
		double &	tuned_accuracy
	)

protected

Get the optimal alpha and the corresponding computation time for a fixed mesh.

_cao should contain an initial guess, which is then adapted by stepping up and down.

Parameters

[in]	mesh	number of mesh points per coordinate direction (>0), in real space.
[in,out]	tuned_cao	initial guess for the

charge assignment order ([0,7]).

Parameters

[out]	tuned_r_cut_iL	cutoff radius for real space electrostatics (>0), rescaled to `r_cut_iL` = `r_cut` * `box_l_i`.
[out]	tuned_alpha_L	Ewald splitting parameter (0.
[out]	tuned_accuracy	accuracy of the actual parameter set.

Returns: The integration time in case of success, otherwise -P3M_TUNE_CAO_TOO_LARGE

Definition at line 223 of file TuningAlgorithm.cpp.

References cao_max, cao_min, get_mc_time(), P3M_TUNE_CAO_TOO_LARGE, and time_granularity.

Referenced by DipolarTuningAlgorithm< FloatType, Architecture >::get_time().

◆ get_mc_time()

double TuningAlgorithm::get_mc_time	(	Utils::Vector3i const &	mesh,
		int	cao,
		double &	tuned_r_cut_iL,
		double &	tuned_alpha_L,
		double &	tuned_accuracy
	)

protected

Get the optimal alpha and the corresponding computation time for a fixed mesh and cao.

The tuned_r_cut_iL is determined via a simple bisection.

Parameters

[in]	mesh	number of mesh points per coordinate direction (>0), in real space.
[in]	cao	charge assignment order ([0,7]).
[in,out]	tuned_r_cut_iL	cutoff radius for real space electrostatics (>0), rescaled to `r_cut_iL` = `r_cut` * `box_l_i`.
[in,out]	tuned_alpha_L	Ewald splitting parameter (0.
[in,out]	tuned_accuracy	accuracy of the actual parameter set.

Returns: The integration time in case of success, otherwise -P3M_TUNE_ACCURACY_TOO_LARGE, -P3M_TUNE_FFT_MESH_SIZE, -P3M_TUNE_CAO_TOO_LARGE, or -P3M_TUNE_ELC_GAP_SIZE

Definition at line 118 of file TuningAlgorithm.cpp.

References P3MParameters::accuracy, benchmark_integration_step(), System::System::box_geo, calculate_accuracy(), System::System::cell_structure, commit(), fft_decomposition_veto(), get_params(), Utils::hadamard_division(), increment_n_trials(), layer_correction_veto_r_cut(), System::System::local_geo, m_logger, m_r_cut_iL_max, m_r_cut_iL_min, m_system, on_solver_change(), P3M_RCUT_PREC, P3M_TUNE_ACCURACY_TOO_LARGE, P3M_TUNE_CAO_TOO_LARGE, P3M_TUNE_ELC_GAP_SIZE, and P3M_TUNE_FFT_MESH_SIZE.

Referenced by get_m_time().

◆ get_n_trials()

auto TuningAlgorithm::get_n_trials ( )

inlineprotected

Definition at line 184 of file TuningAlgorithm.hpp.

Referenced by DipolarTuningAlgorithm< FloatType, Architecture >::get_time().

◆ get_params()

virtual P3MParameters & TuningAlgorithm::get_params ( )

pure virtual

Get the P3M parameters.

Implemented in CoulombTuningAlgorithm< FloatType, Architecture >, and DipolarTuningAlgorithm< FloatType, Architecture >.

Referenced by commit(), determine_cao_limits(), determine_r_cut_limits(), get_mc_time(), and tune().

◆ get_time()

virtual TuningAlgorithm::Parameters TuningAlgorithm::get_time ( )

pure virtual

Tuning loop entry point.

Implemented in CoulombTuningAlgorithm< FloatType, Architecture >, and DipolarTuningAlgorithm< FloatType, Architecture >.

Referenced by tune().

◆ increment_n_trials()

void TuningAlgorithm::increment_n_trials ( )

inlineprotected

Definition at line 185 of file TuningAlgorithm.hpp.

Referenced by get_mc_time().

◆ layer_correction_veto_r_cut()

virtual std::optional< std::string > TuningAlgorithm::layer_correction_veto_r_cut ( double r_cut ) const

pure virtual

Veto real-space cutoffs larger than the layer correction gap.

Implemented in CoulombTuningAlgorithm< FloatType, Architecture >, and DipolarTuningAlgorithm< FloatType, Architecture >.

Referenced by get_mc_time().

◆ on_solver_change()

virtual void TuningAlgorithm::on_solver_change ( ) const

pure virtual

Re-initialize the currently active solver.

Implemented in CoulombTuningAlgorithm< FloatType, Architecture >, and DipolarTuningAlgorithm< FloatType, Architecture >.

Referenced by get_mc_time().

◆ reset_n_trials()

void TuningAlgorithm::reset_n_trials ( )

inlineprotected

Definition at line 186 of file TuningAlgorithm.hpp.

Referenced by DipolarTuningAlgorithm< FloatType, Architecture >::get_time().

◆ setup_logger()

virtual void TuningAlgorithm::setup_logger ( bool verbose )

pure virtual

Configure the logger.

Implemented in CoulombTuningAlgorithm< FloatType, Architecture >, and DipolarTuningAlgorithm< FloatType, Architecture >.

◆ tune()

void TuningAlgorithm::tune ( )

inline

Definition at line 160 of file TuningAlgorithm.hpp.

References P3MParameters::accuracy, commit(), get_params(), get_time(), m_logger, time_sentinel, and P3MParameters::tuning.

Referenced by CoulombP3MImpl< FloatType, Architecture >::tune(), and DipolarP3MImpl< FloatType, Architecture >::tune().

Member Data Documentation

◆ cao_best

int TuningAlgorithm::cao_best = -1

protected

Definition at line 75 of file TuningAlgorithm.hpp.

Referenced by determine_cao_limits(), and DipolarTuningAlgorithm< FloatType, Architecture >::get_time().

◆ cao_max

int TuningAlgorithm::cao_max = -1

protected

Definition at line 75 of file TuningAlgorithm.hpp.

Referenced by determine_cao_limits(), and get_m_time().

◆ cao_min

int TuningAlgorithm::cao_min = -1

protected

Definition at line 75 of file TuningAlgorithm.hpp.

Referenced by determine_cao_limits(), and get_m_time().

◆ m_logger

std::unique_ptr<TuningLogger> TuningAlgorithm::m_logger = nullptr

protected

Definition at line 74 of file TuningAlgorithm.hpp.

Referenced by determine_cao_limits(), DipolarTuningAlgorithm< FloatType, Architecture >::determine_mesh_limits(), determine_r_cut_limits(), get_mc_time(), CoulombTuningAlgorithm< FloatType, Architecture >::setup_logger(), DipolarTuningAlgorithm< FloatType, Architecture >::setup_logger(), and tune().

◆ m_prefactor

double TuningAlgorithm::m_prefactor

protected

Definition at line 73 of file TuningAlgorithm.hpp.

Referenced by CoulombTuningAlgorithm< FloatType, Architecture >::setup_logger(), and DipolarTuningAlgorithm< FloatType, Architecture >::setup_logger().

◆ m_r_cut_iL_max

double TuningAlgorithm::m_r_cut_iL_max = -1.

protected

Definition at line 76 of file TuningAlgorithm.hpp.

Referenced by determine_r_cut_limits(), get_mc_time(), and DipolarTuningAlgorithm< FloatType, Architecture >::get_time().

◆ m_r_cut_iL_min

double TuningAlgorithm::m_r_cut_iL_min = -1.

protected

Definition at line 76 of file TuningAlgorithm.hpp.

Referenced by determine_r_cut_limits(), and get_mc_time().

◆ m_system

System::System& TuningAlgorithm::m_system

protected

Definition at line 66 of file TuningAlgorithm.hpp.

Referenced by DipolarTuningAlgorithm< FloatType, Architecture >::calculate_accuracy(), commit(), determine_r_cut_limits(), get_mc_time(), CoulombTuningAlgorithm< FloatType, Architecture >::layer_correction_veto_r_cut(), CoulombTuningAlgorithm< FloatType, Architecture >::on_solver_change(), DipolarTuningAlgorithm< FloatType, Architecture >::on_solver_change(), CoulombTuningAlgorithm< FloatType, Architecture >::setup_logger(), and DipolarTuningAlgorithm< FloatType, Architecture >::setup_logger().

◆ max_n_consecutive_trials

auto constexpr TuningAlgorithm::max_n_consecutive_trials = 20

staticconstexprprotected

Maximal number of consecutive trials that don't improve runtime.

Tuning halts when this threshold is reached.

Definition at line 88 of file TuningAlgorithm.hpp.

Referenced by DipolarTuningAlgorithm< FloatType, Architecture >::get_time().

◆ time_granularity

auto constexpr TuningAlgorithm::time_granularity = 2.

staticconstexprprotected

Granularity of the time measurement (milliseconds).

Tuning halts when the runtime is larger than the best time plus this value.

Definition at line 82 of file TuningAlgorithm.hpp.

Referenced by get_m_time(), and DipolarTuningAlgorithm< FloatType, Architecture >::get_time().

◆ time_sentinel

auto constexpr TuningAlgorithm::time_sentinel = std::numeric_limits<double>::max()

staticconstexprprotected

Value for invalid time measurements.

Definition at line 91 of file TuningAlgorithm.hpp.

Referenced by DipolarTuningAlgorithm< FloatType, Architecture >::get_time(), and tune().

The documentation for this class was generated from the following files:

Classes

Public Member Functions

Protected Member Functions

Protected Attributes

Static Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ TuningAlgorithm()

◆ ~TuningAlgorithm()

Member Function Documentation

◆ calculate_accuracy()

◆ commit()

◆ determine_cao_limits()

◆ determine_mesh_limits()

◆ determine_r_cut_limits()

◆ fft_decomposition_veto()

◆ get_m_time()

◆ get_mc_time()

◆ get_n_trials()

◆ get_params()

◆ get_time()

◆ increment_n_trials()

◆ layer_correction_veto_r_cut()

◆ on_solver_change()

◆ reset_n_trials()

◆ setup_logger()

◆ tune()

Member Data Documentation

◆ cao_best

◆ cao_max

◆ cao_min

◆ m_logger

◆ m_prefactor

◆ m_r_cut_iL_max

◆ m_r_cut_iL_min

◆ m_system

◆ max_n_consecutive_trials

◆ time_granularity

◆ time_sentinel