#include <davidson.hpp>

Inheritance diagram for q3::PlanewaveDavidsonDiagonalizer:

Public Member Functions
	PlanewaveDavidsonDiagonalizer (const execution::ExecutionPolicy &execution_policy, execution::gpu::blas::GPUBLAS &gpu_blas, gpu::GPUFFTGrid *gpu_fft_grid, execution::gpu::lapack::GPUEigensolver &gpu_eigensolver, std::shared_ptr< execution::gpu::device::GPUDeviceBackend > device_backend)

	PlanewaveDavidsonDiagonalizer (const execution::ExecutionPolicy &execution_policy=default_serial_execution_policy, PlanewaveHamiltonianApplicator hamiltonian_applicator=nullptr, PlanewaveSubspaceManager subspace_manager=nullptr)

void	invalidate_smooth_potential_cache () override

std::string	name () const override

int	number_of_unconverged_bands () const override

void	release_workspace () override

void	set_basis_expansion_factor (int factor) override

void	set_cached_hubbard_projectors (int kpoint_index, const q3::Matrix< Complex > *projectors) override

void	set_cached_projectors (int kpoint_index, const q3::Matrix< Complex > *projectors) override

void	set_hubbard_d_matrix (const std::vector< Complex > *dense_d_matrix_column_major) override

void	set_maximum_iterations (int maximum_iterations) override

void	set_tolerance (Real tolerance) override

void	set_use_mixed_precision (bool use_mixed_precision)

void	solve (int kpoint_index_state, int kpoint_index, int number_of_plane_waves, int number_of_bands, const std::vector< int > &g_vector_indices, const std::vector< Real > &kinetic_energy_squared, const std::vector< Real > &real_space_potential, const Crystal &crystal_structure, const KPointsGrid &kpoint_grid, const PlanewaveBasis &planewave_basis, const NormConservingPlanewaveKohnShamQuantumSystem &pseudopotential_quantum_system, PlanewaveKohnShamQuantumSystem &quantum_system) override

Public Member Functions inherited from q3::PlanewaveDiagonalizer
	~PlanewaveDiagonalizer () override=default

Public Member Functions inherited from q3::Diagonalizer
virtual	~Diagonalizer ()=default

virtual void	set_first_iteration (bool)

Static Public Attributes
static const execution::SerialExecutionPolicy	default_serial_execution_policy {}

Detailed Description

Davidson iterative diagonalizer. Finds only the lowest number_of_bands eigenvalues in O(number_of_plane_waves * number_of_bands) per iteration, vs O(number_of_plane_waves³) for direct.

Definition at line 351 of file davidson.hpp.

Constructor & Destructor Documentation

◆ PlanewaveDavidsonDiagonalizer() [1/2]

q3::PlanewaveDavidsonDiagonalizer::PlanewaveDavidsonDiagonalizer	(	const execution::ExecutionPolicy &	execution_policy = `default_serial_execution_policy`,
		PlanewaveHamiltonianApplicator *	hamiltonian_applicator = `nullptr`,
		PlanewaveSubspaceManager *	subspace_manager = `nullptr`
	)

inlineexplicit

Construct the Davidson diagonalizer with an execution policy for band-group vector operations (dot_product, norm). Defaults to a serial (BLAS-only) policy if none is provided, preserving compatibility with existing call sites.

Optionally accepts a PlanewaveHamiltonianApplicator and/or PlanewaveSubspaceManager for delegation. If not provided, internal serial implementations are created on first solve() call.

Definition at line 365 of file davidson.hpp.

◆ PlanewaveDavidsonDiagonalizer() [2/2]

q3::PlanewaveDavidsonDiagonalizer::PlanewaveDavidsonDiagonalizer	(	const execution::ExecutionPolicy &	execution_policy,
		execution::gpu::blas::GPUBLAS &	gpu_blas,
		gpu::GPUFFTGrid *	gpu_fft_grid,
		execution::gpu::lapack::GPUEigensolver &	gpu_eigensolver,
		std::shared_ptr< execution::gpu::device::GPUDeviceBackend >	device_backend
	)

inline

Construct a GPU-accelerated Davidson diagonalizer.

Stores GPU context (BLAS, FFT, device backend) for lazy construction of GPUPlanewaveHamiltonianApplicator and GPUPlanewaveSubspaceManager on first solve() call, when the Crystal/PlanewaveBasis are available.

Parameters

execution_policy	Execution policy for band-group operations.
gpu_blas	GPU BLAS wrapper. Must outlive this object.
gpu_fft_grid	GPU FFT grid. Must outlive this object.
gpu_eigensolver	GPU eigensolver. Must outlive this object.
device_backend	GPU device backend (shared ownership).

Definition at line 392 of file davidson.hpp.

Member Function Documentation

◆ invalidate_smooth_potential_cache()

void q3::PlanewaveDavidsonDiagonalizer::invalidate_smooth_potential_cache ( )

inlineoverridevirtual

Invalidate the cached smooth grid potential. No-op if the applicator has not been created yet (nothing cached).

Reimplemented from q3::PlanewaveDiagonalizer.

Definition at line 474 of file davidson.hpp.

References q3::PlanewaveHamiltonianApplicator::invalidate_smooth_potential_cache().

◆ name()

std::string q3::PlanewaveDavidsonDiagonalizer::name ( ) const

inlineoverridevirtual

Implements q3::Diagonalizer.

Definition at line 996 of file davidson.hpp.

◆ number_of_unconverged_bands()

int q3::PlanewaveDavidsonDiagonalizer::number_of_unconverged_bands ( ) const

inlineoverridevirtual

How many bands the last solve() returned without meeting its tolerance: zero when it converged, and always zero for a solver that is not iterative.

Reimplemented from q3::PlanewaveDiagonalizer.

Definition at line 432 of file davidson.hpp.

◆ release_workspace()

void q3::PlanewaveDavidsonDiagonalizer::release_workspace ( )

inlineoverridevirtual

Free the working arrays once nothing will call solve() again in this run: the eigenvectors live in the quantum system and survive, the working basis and its Hamiltonian image do not need to. A later solve() allocates afresh. No-op for solvers without a persistent workspace.

Reimplemented from q3::PlanewaveDiagonalizer.

Definition at line 436 of file davidson.hpp.

◆ set_basis_expansion_factor()

void q3::PlanewaveDavidsonDiagonalizer::set_basis_expansion_factor ( int factor )

inlineoverridevirtual

Set pre-computed KB projectors for a specific k-point. No-op if the applicator has not been created yet (before first solve()).

Reimplemented from q3::PlanewaveDiagonalizer.

Definition at line 418 of file davidson.hpp.

◆ set_cached_hubbard_projectors()

void q3::PlanewaveDavidsonDiagonalizer::set_cached_hubbard_projectors	(	int	kpoint_index,
		const q3::Matrix< Complex > *	projectors
	)

inlineoverridevirtual

Set pre-computed DFT+U (Hubbard) atomic-orbital projectors for a specific k-point. No-op if the applicator has not been created yet (before first solve()).

Reimplemented from q3::PlanewaveDiagonalizer.

Definition at line 451 of file davidson.hpp.

References q3::PlanewaveHamiltonianApplicator::set_cached_hubbard_projectors().

◆ set_cached_projectors()

void q3::PlanewaveDavidsonDiagonalizer::set_cached_projectors	(	int	,
		const q3::Matrix< Complex > *
	)

inlineoverridevirtual

Set pre-computed KB projectors for a specific k-point. If set, the solver may use these instead of recomputing them. No-op by default; overridden by solvers that support caching.

Reimplemented from q3::PlanewaveDiagonalizer.

Definition at line 440 of file davidson.hpp.

References q3::PlanewaveHamiltonianApplicator::set_cached_projectors().

◆ set_hubbard_d_matrix()

void q3::PlanewaveDavidsonDiagonalizer::set_hubbard_d_matrix ( const std::vector< Complex > * dense_d_matrix_column_major )

inlineoverridevirtual

Set the dense block-diagonal Hubbard D-matrix (refreshed once per SCF iteration). No-op if the applicator has not been created yet (before first solve()).

Reimplemented from q3::PlanewaveDiagonalizer.

Definition at line 463 of file davidson.hpp.

References q3::PlanewaveHamiltonianApplicator::set_hubbard_d_matrix().

◆ set_maximum_iterations()

void q3::PlanewaveDavidsonDiagonalizer::set_maximum_iterations ( int maximum_iterations )

inlineoverridevirtual

Drop the working basis, its Hamiltonian image and the subspace: the largest thing the solver owns, and dead once the SCF has converged.

Reimplemented from q3::PlanewaveDiagonalizer.

Definition at line 428 of file davidson.hpp.

◆ set_tolerance()

void q3::PlanewaveDavidsonDiagonalizer::set_tolerance ( Real )

inlineoverridevirtual

Set the convergence tolerance for iterative eigensolvers. No-op for direct solvers; overridden by Davidson and similar.

Reimplemented from q3::Diagonalizer.

Definition at line 1000 of file davidson.hpp.

◆ set_use_mixed_precision()

void q3::PlanewaveDavidsonDiagonalizer::set_use_mixed_precision ( bool use_mixed_precision )

inline

Enable mixed-precision mode for GPU Davidson.

When enabled, the local potential FFT section of H|ψ⟩ runs in FP32 while kinetic energy and nonlocal projectors remain FP64.

Definition at line 410 of file davidson.hpp.

◆ solve()

void q3::PlanewaveDavidsonDiagonalizer::solve	(	int	kpoint_index_state,
		int	kpoint_index,
		int	number_of_plane_waves,
		int	number_of_bands,
		const std::vector< int > &	g_vector_indices,
		const std::vector< Real > &	kinetic_energy_squared,
		const std::vector< Real > &	real_space_potential,
		const Crystal &	crystal_structure,
		const KPointsGrid &	kpoint_grid,
		const PlanewaveBasis &	planewave_basis,
		const NormConservingPlanewaveKohnShamQuantumSystem &	pseudopotential_quantum_system,
		PlanewaveKohnShamQuantumSystem &	quantum_system
	)

inlineoverridevirtual

Solve the KS eigenvalue problem at one k-point. For NC-PP: H·ψ = ε·ψ (standard eigenvalue problem) For US-PP: H·ψ = ε·S·ψ (generalized eigenvalue problem)

Which spin channel this call belongs to.

The caller flattens (spin, k-point) into kpoint_index_state, so the channel is the quotient. The applicator needs it because the nonlocal D is per channel for ultrasoft and PAW; real_space_potential below is already this channel's, and the two have to agree.

Under a reciprocal-space distribution one rank holds G = 0, and the manager's gamma corrections must know whether this is that rank.

Only the size needs agreeing here. The collapse itself now forms each band range on one rank and allgathers, so the collapsed basis is already identical everywhere; republishing it would move three full arrays for nothing.

No Ritz vectors are formed. Convergence is decided on the eigenvalues, and the correction vectors below are built straight from the subspace coefficients into the free basis columns.

Expand subspace with preconditioned residuals if not converged

Compute residuals and apply preconditioner.

Build the preconditioned corrections directly into the columns the basis is about to grow into, compacted over the unconverged bands. Nothing is staged through a residual array on the way.

Only the size needs agreeing here. The collapse itself now forms each band range on one rank and allgathers, so the collapsed basis is already identical everywhere; republishing it would move three full arrays for nothing.

Implements q3::PlanewaveDiagonalizer.

Definition at line 480 of file davidson.hpp.

References q3::NormConservingPlanewaveBasis::beta_function_table(), q3::execution::ExecutionPolicy::broadcast_across_bands(), q3::ProfilingCounters::DavidsonTimingAccumulator::collapse_time_ms, q3::compute_kleinman_bylander_projectors(), q3::create_overlap_operator(), q3::ProfilingCounters::davidson_iteration_count, q3::ProfilingCounters::davidson_timing, q3::log::debug(), q3::ProfilingCounters::DavidsonTimingAccumulator::diagonalize_time_ms, q3::PlanewaveKohnShamQuantumSystem::eigenvalue(), q3::math::tolerances::empty_band_tolerance_floor, q3::math::tolerances::empty_band_tolerance_relaxation_factor, q3::execution::ExecutionPolicy::global_number_of_elements(), q3::ProfilingCounters::DavidsonTimingAccumulator::hamiltonian_application_time_ms, q3::ProfilingCounters::instance(), q3::KPointsGrid::number_of_kpoints(), q3::PlanewaveKohnShamQuantumSystem::number_of_occupied_bands_per_spin(), q3::ProfilingCounters::DavidsonTimingAccumulator::orthogonalize_time_ms, q3::NormConservingPlanewaveKohnShamQuantumSystem::pseudopotential_basis(), q3::NormConservingPlanewaveBasis::pseudopotential_set(), q3::ProfilingCounters::DavidsonTimingAccumulator::residual_time_ms, q3::ProfilingCounters::DavidsonTimingAccumulator::ritz_vector_time_ms, q3::ProfilingCounters::DavidsonTimingAccumulator::seed_time_ms, q3::ProfilingCounters::DavidsonTimingAccumulator::subspace_update_time_ms, q3::NormConservingPlanewaveBasis::total_number_of_projectors(), and q3::log::warn().

Member Data Documentation

◆ default_serial_execution_policy

const execution::SerialExecutionPolicy q3::PlanewaveDavidsonDiagonalizer::default_serial_execution_policy {}

inlinestatic

Definition at line 354 of file davidson.hpp.

Referenced by q3::cli::run_gpu_entry_point().

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

src/layer_4_computation/diagonalizer/planewave/davidson.hpp

Public Member Functions

Static Public Attributes

Detailed Description

Constructor & Destructor Documentation

◆ PlanewaveDavidsonDiagonalizer() [1/2]

◆ PlanewaveDavidsonDiagonalizer() [2/2]

Member Function Documentation

◆ invalidate_smooth_potential_cache()

◆ name()

◆ number_of_unconverged_bands()

◆ release_workspace()

◆ set_basis_expansion_factor()

◆ set_cached_hubbard_projectors()

◆ set_cached_projectors()

◆ set_hubbard_d_matrix()

◆ set_maximum_iterations()

◆ set_tolerance()

◆ set_use_mixed_precision()

◆ solve()

Member Data Documentation

◆ default_serial_execution_policy