#include <gpu.hpp>

Inheritance diagram for q3::GPUPlanewaveHamiltonianApplicator:

Public Member Functions
	GPUPlanewaveHamiltonianApplicator (const Crystal &crystal_structure, const PlanewaveBasis &planewave_basis, const KPointsGrid &kpoint_grid, const NormConservingPlanewaveKohnShamQuantumSystem &pseudopotential_quantum_system, execution::gpu::blas::GPUBLAS &gpu_blas, gpu::GPUFFTGrid *gpu_fft_grid, std::shared_ptr< execution::gpu::device::GPUDeviceBackend > device_backend)

void	apply_batch (const Complex input_wavefunctions, Complex output_hamiltonian_wavefunctions, int number_of_vectors) override

void	apply_batch_single_precision (const ComplexSingle input_wavefunctions, ComplexSingle output_hamiltonian_wavefunctions, int number_of_vectors)

const std::vector< Real > &	cached_scaled_potential () const

const std::vector< int > &	g_to_fft_mapping () const

const q3::Matrix< Complex > *	get_cached_projectors (int) const override

void	invalidate_smooth_potential_cache () override

bool	is_gamma_only_current () const override

bool	is_smooth_potential_cache_valid () const override

const std::vector< int > &	minus_g_to_fft_mapping () const

void	prepare_kpoint (int kpoint_index, int number_of_plane_waves, const std::vector< int > &g_vector_indices, const std::vector< Real > &kinetic_energy_squared) override

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

void	set_use_mixed_precision (bool use_mixed_precision)

void	update_potential (const std::vector< Real > &real_space_potential) override

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

virtual bool	gamma_origin_is_local_current () const

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

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

void	set_spin_channel (int spin_channel)

int	spin_channel () const

Additional Inherited Members
Protected Attributes inherited from q3::PlanewaveHamiltonianApplicator
int	spin_channel_ {0}
	Spin channel the current potential and D belong to; 0 when unpolarized.

Detailed Description

Definition at line 51 of file gpu.hpp.

Constructor & Destructor Documentation

◆ GPUPlanewaveHamiltonianApplicator()

q3::GPUPlanewaveHamiltonianApplicator::GPUPlanewaveHamiltonianApplicator	(	const Crystal &	crystal_structure,
		const PlanewaveBasis &	planewave_basis,
		const KPointsGrid &	kpoint_grid,
		const NormConservingPlanewaveKohnShamQuantumSystem &	pseudopotential_quantum_system,
		execution::gpu::blas::GPUBLAS &	gpu_blas,
		gpu::GPUFFTGrid *	gpu_fft_grid,
		std::shared_ptr< execution::gpu::device::GPUDeviceBackend >	device_backend
	)

inline

Construct with run-constant planewave context and GPU math wrappers.

All references and the shared_ptr must outlive this object.

Parameters

crystal_structure	Crystal geometry.
planewave_basis	Planewave basis (G-vectors, FFT grids).
kpoint_grid	K-point grid for Brillouin zone sampling.
pseudopotential_quantum_system	Pseudopotential data (KB projectors, D-matrix).
gpu_blas	GPU BLAS wrapper for matrix/vector operations.
gpu_fft_grid	GPU FFT grid for potential application.
device_backend	GPU device backend for memory allocation.

Initialize the GPU FFT grid with the wave (smooth) grid dimensions so that cuFFT plans are created before any FFT calls.

Definition at line 66 of file gpu.hpp.

References q3::PlanewaveBasis::fft_wave(), and q3::FFTGrid::init().

Member Function Documentation

◆ apply_batch()

void q3::GPUPlanewaveHamiltonianApplicator::apply_batch	(	const Complex *	input_wavefunctions,
		Complex *	output_hamiltonian_wavefunctions,
		int	number_of_vectors
	)

inlineoverridevirtual

Apply H|ψ⟩ in double precision (FP64) to a batch of vectors.

Uses GPU BLAS for KB projector application (3× ZGEMM) and CPU-side FFT for the local potential application. The kinetic energy is applied element-wise on the host (unified memory).

All input/output pointers must come from GPUDeviceBuffer (unified memory).

Implements q3::PlanewaveHamiltonianApplicator.

Definition at line 258 of file gpu.hpp.

References q3::execution::gpu::device::GPUDeviceBuffer< T >::data(), q3::PlanewaveBasis::fft_wave(), q3::diagonalizer::planewave::gpu::launch_apply_kinetic_energy(), q3::FFTGrid::local_grid_points(), and q3::execution::gpu::device::GPUDeviceBuffer< T >::size().

◆ apply_batch_single_precision()

void q3::GPUPlanewaveHamiltonianApplicator::apply_batch_single_precision	(	const ComplexSingle *	input_wavefunctions,
		ComplexSingle *	output_hamiltonian_wavefunctions,
		int	number_of_vectors
	)

inline

Apply H|ψ⟩ in single precision (FP32) to a batch of vectors.

Identical algorithm to apply_batch() but operates on ComplexSingle arrays and uses FP32 BLAS/FFT calls. Used by the mixed-precision Davidson loop for bulk iterations before switchover.

Local potential via batched FFT in FP32. Scatter all bands → batch inverse FFT → V·ψ for all → batch forward FFT → gather all. This replaces 2×N single FFT calls with 2 batched kernel launches.

Definition at line 313 of file gpu.hpp.

References q3::execution::gpu::device::GPUDeviceBuffer< T >::data(), q3::PlanewaveBasis::fft_wave(), q3::gpu::GPUFFTGrid::forward_single_batch(), q3::gpu::GPUFFTGrid::inverse_raw_single_batch(), q3::PlanewaveBasis::is_gamma_only(), q3::diagonalizer::planewave::gpu::launch_apply_kinetic_energy_single(), q3::diagonalizer::planewave::gpu::launch_gather_fft_grid_to_g_single(), q3::diagonalizer::planewave::gpu::launch_multiply_potential_single(), q3::diagonalizer::planewave::gpu::launch_scatter_g_to_fft_grid_single(), q3::FFTGrid::local_grid_points(), and q3::execution::gpu::device::GPUDeviceBuffer< T >::size().

◆ cached_scaled_potential()

const std::vector< Real > & q3::GPUPlanewaveHamiltonianApplicator::cached_scaled_potential ( ) const

inline

Access the cached scaled potential (for serial fallback or debugging).

Definition at line 406 of file gpu.hpp.

◆ g_to_fft_mapping()

const std::vector< int > & q3::GPUPlanewaveHamiltonianApplicator::g_to_fft_mapping ( ) const

inline

Access the G-to-FFT mapping for the current k-point.

Definition at line 413 of file gpu.hpp.

◆ get_cached_projectors()

const q3::Matrix< Complex > * q3::GPUPlanewaveHamiltonianApplicator::get_cached_projectors ( int ) const

inlineoverridevirtual

Returns nullptr — GPU applicator manages projectors internally.

Reimplemented from q3::PlanewaveHamiltonianApplicator.

Definition at line 444 of file gpu.hpp.

◆ invalidate_smooth_potential_cache()

void q3::GPUPlanewaveHamiltonianApplicator::invalidate_smooth_potential_cache ( )

inlineoverridevirtual

Invalidate any cached smooth potential. Called when the effective potential changes (e.g., at start of each SCF iteration).

Implements q3::PlanewaveHamiltonianApplicator.

Definition at line 391 of file gpu.hpp.

◆ is_gamma_only_current()

bool q3::GPUPlanewaveHamiltonianApplicator::is_gamma_only_current ( ) const

inlineoverridevirtual

Report whether the current k-point is gamma-only.

This affects dot product computations in the subspace manager: gamma-only uses 2×Re(z·z*) instead of complex zdotc.

Reimplemented from q3::PlanewaveHamiltonianApplicator.

Definition at line 430 of file gpu.hpp.

References q3::PlanewaveBasis::is_gamma_only().

◆ is_smooth_potential_cache_valid()

bool q3::GPUPlanewaveHamiltonianApplicator::is_smooth_potential_cache_valid ( ) const

inlineoverridevirtual

Whether the smooth potential cache is currently valid.

Reimplemented from q3::PlanewaveHamiltonianApplicator.

Definition at line 399 of file gpu.hpp.

◆ minus_g_to_fft_mapping()

const std::vector< int > & q3::GPUPlanewaveHamiltonianApplicator::minus_g_to_fft_mapping ( ) const

inline

Access the minus-G-to-FFT mapping (for gamma-only scatter/gather).

Definition at line 420 of file gpu.hpp.

◆ prepare_kpoint()

void q3::GPUPlanewaveHamiltonianApplicator::prepare_kpoint	(	int	kpoint_index,
		int	number_of_plane_waves,
		const std::vector< int > &	g_vector_indices,
		const std::vector< Real > &	kinetic_energy_squared
	)

inlineoverridevirtual

Configure for a specific k-point.

Copies G-to-FFT mapping, kinetic energy factors, and KB projectors to GPUDeviceBuffers. The KB projectors are computed on CPU and then transferred to the GPU.

Implements q3::PlanewaveHamiltonianApplicator.

Definition at line 167 of file gpu.hpp.

References q3::NormConservingPlanewaveBasis::beta_function_table(), q3::precision::cast_double_to_single(), q3::precision::cast_real_double_to_single(), q3::compute_kleinman_bylander_projectors(), q3::Matrix< T >::data(), q3::execution::gpu::device::GPUDeviceBuffer< T >::data(), q3::Matrix< T >::empty(), q3::PlanewaveBasis::is_gamma_only(), q3::NormConservingPlanewaveKohnShamQuantumSystem::pseudopotential_basis(), q3::NormConservingPlanewaveBasis::pseudopotential_set(), q3::Matrix< T >::rows(), and q3::NormConservingPlanewaveBasis::total_number_of_projectors().

◆ set_cached_projectors()

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

inlineoverridevirtual

No-op: GPU applicator computes projectors internally in prepare_kpoint().

Reimplemented from q3::PlanewaveHamiltonianApplicator.

Definition at line 437 of file gpu.hpp.

◆ set_use_mixed_precision()

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

inline

Enable or disable mixed-precision mode.

When enabled, apply_batch() runs the local potential FFT section in FP32 (scatter FP64→FP32, FFT, V*ψ, FFT, gather FP32→FP64). Kinetic energy and nonlocal projectors remain in FP64. This gives most of the GPU FP32 speedup without restructuring the Davidson loop.

Definition at line 115 of file gpu.hpp.

References q3::log::info().

◆ update_potential()

void q3::GPUPlanewaveHamiltonianApplicator::update_potential ( const std::vector< Real > & real_space_potential )

inlineoverridevirtual

Update the effective potential for a new SCF iteration.

Computes the smooth-grid potential on CPU (same as serial), then copies the result to a GPUDeviceBuffer for GPU-side V(r)·ψ(r) multiplication.

Implements q3::PlanewaveHamiltonianApplicator.

Definition at line 129 of file gpu.hpp.

References q3::precision::cast_real_double_to_single(), q3::execution::gpu::device::GPUDeviceBuffer< T >::data(), q3::PlanewaveBasis::fft_wave(), q3::FFTGrid::local_grid_points(), and q3::FFTGrid::total_grid_points().

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

src/layer_4_computation/diagonalizer/planewave/internal/hamiltonian_applicator/gpu.hpp

Public Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ GPUPlanewaveHamiltonianApplicator()

Member Function Documentation

◆ apply_batch()

◆ apply_batch_single_precision()

◆ cached_scaled_potential()

◆ g_to_fft_mapping()

◆ get_cached_projectors()

◆ invalidate_smooth_potential_cache()

◆ is_gamma_only_current()

◆ is_smooth_potential_cache_valid()

◆ minus_g_to_fft_mapping()

◆ prepare_kpoint()

◆ set_cached_projectors()

◆ set_use_mixed_precision()

◆ update_potential()