#include <init.hpp>

Inheritance diagram for q3::gpu::GPUFFTGrid:

Public Member Functions
	GPUFFTGrid (std::shared_ptr< execution::gpu::device::GPUDeviceBackend > device_backend)

	~GPUFFTGrid () override=default

execution::gpu::device::GPUDeviceBackend &	device_backend ()

const execution::gpu::device::GPUDeviceBackend &	device_backend () const

virtual void	forward_single (ComplexSingle *data) const =0

virtual void	forward_single_batch (ComplexSingle *data, int batch_size) const

virtual void	forward_transform (Complex *data) const

virtual void	forward_transform_batch (Complex *data, int batch_size) const

virtual void	inverse_raw_single (ComplexSingle *data) const =0

virtual void	inverse_raw_single_batch (ComplexSingle *data, int batch_size) const

virtual void	inverse_transform (Complex *data) const

virtual void	inverse_transform_batch (Complex *data, int batch_size) const

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

virtual void	assign_sticks (const std::vector< int > &)

virtual bool	co_locates_mirrors () const

virtual int	concurrent_transform_slots () const

virtual void	execute_forward_on_buffers () const =0

virtual void	execute_forward_on_slot (int) const
	Forward transform on slot `slot`, the slot-aware `execute_forward_on_buffers()`.

virtual void	execute_inverse_raw_on_buffers () const =0

virtual void	execute_inverse_raw_on_slot (int) const
	Inverse transform on slot `slot`, the slot-aware `execute_inverse_raw_on_buffers()`.

virtual void	forward (Complex *data) const =0

virtual void	forward_batched (Complex *data, int batch_size) const

virtual int	g_vector_to_fft_grid_index (int miller_1, int miller_2, int miller_3) const =0

virtual void	gather_local_g_space (const Complex full_g_space, Complex local_g_space) const

virtual int	global_real_space_index (int local_index) const

virtual int	grid_dimension_1 () const =0

virtual int	grid_dimension_2 () const =0

virtual int	grid_dimension_3 () const =0

virtual void	init (int dimension_1, int dimension_2, int dimension_3, bool use_measured_planning=false)=0

virtual void	initialize_batched_plans (int, bool=false) const

virtual int	initialized_batched_plan_size () const

virtual Complex *	input_buffer () const =0

virtual void	inverse (Complex *data) const =0

virtual void	inverse_raw (Complex *data) const =0

virtual void	inverse_raw_batched (Complex *data, int batch_size) const

virtual void	inverse_raw_on_slot (Complex *data, int) const

virtual int	local_g_space_points () const

virtual int	local_grid_points () const =0

virtual int	local_offset () const =0

virtual Complex *	output_buffer () const =0

virtual bool	owns_stick (int, int) const

virtual void	prepare_forward_from_output () const =0

virtual void	prepare_forward_from_output_on_slot (int) const
	Output-to-input copy on slot `slot`, the slot-aware `prepare_forward_from_output()`.

virtual void	reduce_g_space (Complex *data, int number_of_elements) const =0

virtual void	scatter_local_g_space (const Complex local_g_space, Complex full_g_space) const
	The reverse: place this rank's reciprocal-space points into a full array.

virtual int	slab_dimension_1 () const

virtual Complex *	slot_input_buffer (int) const

virtual Complex *	slot_output_buffer (int) const

virtual int	total_grid_points () const =0

Protected Attributes
std::shared_ptr< execution::gpu::device::GPUDeviceBackend >	device_backend_

Detailed Description

Definition at line 36 of file init.hpp.

Constructor & Destructor Documentation

◆ GPUFFTGrid()

q3::gpu::GPUFFTGrid::GPUFFTGrid ( std::shared_ptr< execution::gpu::device::GPUDeviceBackend > device_backend )

inlineexplicit

Construct a GPU FFT grid backed by the given device.

FFT plans are created during init(), not in the constructor. The device backend is stored for memory management and synchronization.

Parameters

device_backend Shared pointer to the GPU device backend. Must not be null.

Definition at line 48 of file init.hpp.

◆ ~GPUFFTGrid()

q3::gpu::GPUFFTGrid::~GPUFFTGrid ( )

overridedefault

Member Function Documentation

◆ device_backend() [1/2]

execution::gpu::device::GPUDeviceBackend & q3::gpu::GPUFFTGrid::device_backend ( )

inline

Access the underlying GPU device backend.

Definition at line 162 of file init.hpp.

References device_backend_.

◆ device_backend() [2/2]

const execution::gpu::device::GPUDeviceBackend & q3::gpu::GPUFFTGrid::device_backend ( ) const

inline

Access the underlying GPU device backend (const).

Definition at line 169 of file init.hpp.

References device_backend_.

◆ forward_single()

virtual void q3::gpu::GPUFFTGrid::forward_single ( ComplexSingle * data ) const

pure virtual

Forward single-precision FFT: real-space → reciprocal-space (unnormalized).

Parameters

data	In-out: single-precision complex grid data.

Implemented in q3::gpu::CudaGPUFFTGrid, q3::gpu::HIPGPUFFTGrid, q3::gpu::MetalGPUFFTGrid, and q3::gpu::SYCLGPUFFTGrid.

Referenced by forward_single_batch().

◆ forward_single_batch()

virtual void q3::gpu::GPUFFTGrid::forward_single_batch	(	ComplexSingle *	data,
		int	batch_size
	)		const

inlinevirtual

Batched forward single-precision FFT on multiple contiguous grids.

Each grid is total_grid_points() contiguous ComplexSingle values. Default loops over forward_single(); CUDA overrides with cufftPlanMany.

Parameters

data	Pointer to batch_size contiguous FFT grids.
batch_size	Number of grids to transform.

Reimplemented in q3::gpu::CudaGPUFFTGrid, and q3::gpu::HIPGPUFFTGrid.

Definition at line 78 of file init.hpp.

References forward_single(), and q3::FFTGrid::total_grid_points().

Referenced by q3::GPUPlanewaveHamiltonianApplicator::apply_batch_single_precision().

◆ forward_transform()

virtual void q3::gpu::GPUFFTGrid::forward_transform ( Complex * data ) const

inlinevirtual

Forward FFT (FP64). On CUDA, delegates to cuFFT. On Metal, falls back to CPU FFTW since Apple GPUs lack native FP64 compute.

Default implementation delegates to the base FFTGrid (CPU FFTW). Subclasses with native FP64 GPU FFT support should override.

Reimplemented in q3::gpu::CudaGPUFFTGrid, q3::gpu::HIPGPUFFTGrid, and q3::gpu::SYCLGPUFFTGrid.

Definition at line 110 of file init.hpp.

References q3::FFTGrid::forward().

Referenced by forward_transform_batch().

◆ forward_transform_batch()

virtual void q3::gpu::GPUFFTGrid::forward_transform_batch	(	Complex *	data,
		int	batch_size
	)		const

inlinevirtual

Batched forward FP64 FFT on multiple contiguous grids.

Each grid is total_grid_points() contiguous Complex values. Default loops over forward_transform(); CUDA overrides with cufftPlanMany.

Parameters

data	Pointer to batch_size contiguous FFT grids.
batch_size	Number of grids to transform.

Reimplemented in q3::gpu::CudaGPUFFTGrid, and q3::gpu::HIPGPUFFTGrid.

Definition at line 136 of file init.hpp.

References forward_transform(), and q3::FFTGrid::total_grid_points().

◆ inverse_raw_single()

virtual void q3::gpu::GPUFFTGrid::inverse_raw_single ( ComplexSingle * data ) const

pure virtual

Inverse single-precision FFT without normalization (raw backward).

Parameters

data	In-out: single-precision complex grid data.

Implemented in q3::gpu::CudaGPUFFTGrid, q3::gpu::HIPGPUFFTGrid, q3::gpu::MetalGPUFFTGrid, and q3::gpu::SYCLGPUFFTGrid.

Referenced by inverse_raw_single_batch().

◆ inverse_raw_single_batch()

virtual void q3::gpu::GPUFFTGrid::inverse_raw_single_batch	(	ComplexSingle *	data,
		int	batch_size
	)		const

inlinevirtual

Batched inverse single-precision FFT on multiple contiguous grids.

Each grid is total_grid_points() contiguous ComplexSingle values. Default loops over inverse_raw_single(); CUDA overrides with cufftPlanMany.

Parameters

data	Pointer to batch_size contiguous FFT grids.
batch_size	Number of grids to transform.

Reimplemented in q3::gpu::CudaGPUFFTGrid, and q3::gpu::HIPGPUFFTGrid.

Definition at line 94 of file init.hpp.

References inverse_raw_single(), and q3::FFTGrid::total_grid_points().

Referenced by q3::GPUPlanewaveHamiltonianApplicator::apply_batch_single_precision().

◆ inverse_transform()

virtual void q3::gpu::GPUFFTGrid::inverse_transform ( Complex * data ) const

inlinevirtual

Inverse FFT (FP64). On CUDA, delegates to cuFFT. On Metal, falls back to CPU FFTW since Apple GPUs lack native FP64 compute.

Default implementation delegates to the base FFTGrid (CPU FFTW). Subclasses with native FP64 GPU FFT support should override.

Reimplemented in q3::gpu::CudaGPUFFTGrid, q3::gpu::HIPGPUFFTGrid, and q3::gpu::SYCLGPUFFTGrid.

Definition at line 122 of file init.hpp.

References q3::FFTGrid::inverse_raw().

Referenced by inverse_transform_batch().

◆ inverse_transform_batch()

virtual void q3::gpu::GPUFFTGrid::inverse_transform_batch	(	Complex *	data,
		int	batch_size
	)		const

inlinevirtual

Batched inverse FP64 FFT on multiple contiguous grids.

Each grid is total_grid_points() contiguous Complex values. Default loops over inverse_transform(); CUDA overrides with cufftPlanMany.

Parameters

data	Pointer to batch_size contiguous FFT grids.
batch_size	Number of grids to transform.

Reimplemented in q3::gpu::CudaGPUFFTGrid, and q3::gpu::HIPGPUFFTGrid.

Definition at line 152 of file init.hpp.

References inverse_transform(), and q3::FFTGrid::total_grid_points().

Member Data Documentation

◆ device_backend_

std::shared_ptr<execution::gpu::device::GPUDeviceBackend> q3::gpu::GPUFFTGrid::device_backend_

protected

Definition at line 174 of file init.hpp.

Referenced by device_backend(), and device_backend().

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

src/layer_3_representation/basis/planewave/internal/fft/gpu/init.hpp

Public Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ GPUFFTGrid()

◆ ~GPUFFTGrid()

Member Function Documentation

◆ device_backend() [1/2]

◆ device_backend() [2/2]

◆ forward_single()

◆ forward_single_batch()

◆ forward_transform()

◆ forward_transform_batch()

◆ inverse_raw_single()

◆ inverse_raw_single_batch()

◆ inverse_transform()

◆ inverse_transform_batch()

Member Data Documentation

◆ device_backend_