#include <init.hpp>

Inheritance diagram for q3::execution::gpu::blas::GPUBLAS:

Public Member Functions
	GPUBLAS (std::shared_ptr< device::GPUDeviceBackend > device_backend)

virtual	~GPUBLAS ()=default

virtual void	axpy (int number_of_elements, Complex scaling_factor, const Complex vector_x, Complex vector_y) const =0

device::GPUDeviceBackend &	device_backend ()

const device::GPUDeviceBackend &	device_backend () const

virtual void	dgemm (char transpose_a, char transpose_b, int m, int n, int k, Real alpha, const Real matrix_a, int leading_dimension_a, const Real matrix_b, int leading_dimension_b, Real beta, Real *matrix_c, int leading_dimension_c) const =0

virtual Complex	dot_product (int number_of_elements, const Complex vector_x, const Complex vector_y) const =0

virtual ComplexSingle	dot_product_single (int number_of_elements, const ComplexSingle vector_x, const ComplexSingle vector_y) const =0

virtual void	gemm (char transpose_a, char transpose_b, int m, int n, int k, Complex alpha, const Complex matrix_a, int leading_dimension_a, const Complex matrix_b, int leading_dimension_b, Complex beta, Complex *matrix_c, int leading_dimension_c) const =0

virtual void	gemm_single (char transpose_a, char transpose_b, int m, int n, int k, ComplexSingle alpha, const ComplexSingle matrix_a, int leading_dimension_a, const ComplexSingle matrix_b, int leading_dimension_b, ComplexSingle beta, ComplexSingle *matrix_c, int leading_dimension_c) const =0

virtual double	norm (int number_of_elements, const Complex *vector_x) const =0

virtual float	norm_single (int number_of_elements, const ComplexSingle *vector_x) const =0

virtual void	scale (int number_of_elements, Complex scaling_factor, Complex *vector_x) const =0

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

Detailed Description

Definition at line 33 of file init.hpp.

Constructor & Destructor Documentation

◆ GPUBLAS()

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

inlineexplicit

Construct a GPU BLAS wrapper backed by the given device.

The device backend is used for synchronization. The concrete vendor BLAS handle (cuBLAS, Accelerate) is created internally by the vendor-specific subclass constructor.

Parameters

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

Definition at line 45 of file init.hpp.

◆ ~GPUBLAS()

virtual q3::execution::gpu::blas::GPUBLAS::~GPUBLAS ( )

virtualdefault

Member Function Documentation

◆ axpy()

virtual void q3::execution::gpu::blas::GPUBLAS::axpy	(	int	number_of_elements,
		Complex	scaling_factor,
		const Complex *	vector_x,
		Complex *	vector_y
	)		const

pure virtual

Scaled vector addition: vector_y += scaling_factor * vector_x (ZAXPY).

Parameters

number_of_elements	Length of both vectors.
scaling_factor	Complex scaling coefficient.
vector_x	Source vector.
vector_y	Destination vector (modified in-place).

Implemented in q3::execution::gpu::blas::CudaGPUBLAS, q3::execution::gpu::blas::HIPGPUBLAS, and q3::execution::gpu::blas::SYCLGPUBLAS.

Referenced by q3::execution::PlanewaveGPUExecutionPolicy::add_scaled_vector().

◆ device_backend() [1/2]

device::GPUDeviceBackend & q3::execution::gpu::blas::GPUBLAS::device_backend ( )

inline

Access the underlying GPU device backend.

Definition at line 200 of file init.hpp.

References device_backend_.

◆ device_backend() [2/2]

const device::GPUDeviceBackend & q3::execution::gpu::blas::GPUBLAS::device_backend ( ) const

inline

Access the underlying GPU device backend (const).

Definition at line 207 of file init.hpp.

References device_backend_.

◆ dgemm()

virtual void q3::execution::gpu::blas::GPUBLAS::dgemm	(	char	transpose_a,
		char	transpose_b,
		int	m,
		int	n,
		int	k,
		Real	alpha,
		const Real *	matrix_a,
		int	leading_dimension_a,
		const Real *	matrix_b,
		int	leading_dimension_b,
		Real	beta,
		Real *	matrix_c,
		int	leading_dimension_c
	)		const

pure virtual

Real double-precision general matrix-matrix multiply (DGEMM). matrix_c := alpha * op(matrix_a) * op(matrix_b) + beta * matrix_c

Used in gamma-only calculations where wavefunctions are real-valued.

Implemented in q3::execution::gpu::blas::CudaGPUBLAS, q3::execution::gpu::blas::HIPGPUBLAS, and q3::execution::gpu::blas::SYCLGPUBLAS.

Referenced by q3::GPUPlanewaveSubspaceManager::orthogonalize_batch(), and q3::GPUPlanewaveSubspaceManager::update_subspace_matrices().

◆ dot_product()

virtual Complex q3::execution::gpu::blas::GPUBLAS::dot_product	(	int	number_of_elements,
		const Complex *	vector_x,
		const Complex *	vector_y
	)		const

pure virtual

Complex double-precision conjugated dot product: result = x† · y.

Synchronizes before returning the scalar result.

Parameters

number_of_elements	Length of both vectors.
vector_x	Left operand (conjugated).
vector_y	Right operand.

Returns: The complex dot product.

Implemented in q3::execution::gpu::blas::CudaGPUBLAS, q3::execution::gpu::blas::HIPGPUBLAS, and q3::execution::gpu::blas::SYCLGPUBLAS.

Referenced by q3::execution::PlanewaveGPUExecutionPolicy::dot_product(), and q3::cli::validate_gpu_blas().

◆ dot_product_single()

virtual ComplexSingle q3::execution::gpu::blas::GPUBLAS::dot_product_single	(	int	number_of_elements,
		const ComplexSingle *	vector_x,
		const ComplexSingle *	vector_y
	)		const

pure virtual

Complex single-precision conjugated dot product: result = x† · y.

Synchronizes before returning the scalar result.

Parameters

number_of_elements	Length of both vectors.
vector_x	Left operand (conjugated).
vector_y	Right operand.

Returns: The complex single-precision dot product.

Implemented in q3::execution::gpu::blas::CudaGPUBLAS, q3::execution::gpu::blas::HIPGPUBLAS, and q3::execution::gpu::blas::SYCLGPUBLAS.

◆ gemm()

virtual void q3::execution::gpu::blas::GPUBLAS::gemm	(	char	transpose_a,
		char	transpose_b,
		int	m,
		int	n,
		int	k,
		Complex	alpha,
		const Complex *	matrix_a,
		int	leading_dimension_a,
		const Complex *	matrix_b,
		int	leading_dimension_b,
		Complex	beta,
		Complex *	matrix_c,
		int	leading_dimension_c
	)		const

pure virtual

Complex double-precision general matrix-matrix multiply (ZGEMM). matrix_c := alpha * op(matrix_a) * op(matrix_b) + beta * matrix_c

Parameters

transpose_a	'N', 'T', or 'C' for op(matrix_a).
transpose_b	'N', 'T', or 'C' for op(matrix_b).
m	Number of rows of op(matrix_a) and matrix_c.
n	Number of columns of op(matrix_b) and matrix_c.
k	Number of columns of op(matrix_a) / rows of op(matrix_b).
alpha	Scalar multiplier for the product.
matrix_a	Pointer to matrix A data.
leading_dimension_a	Leading dimension of matrix_a.
matrix_b	Pointer to matrix B data.
leading_dimension_b	Leading dimension of matrix_b.
beta	Scalar multiplier for matrix_c.
matrix_c	Pointer to matrix C data (in-out).
leading_dimension_c	Leading dimension of matrix_c.

Implemented in q3::execution::gpu::blas::CudaGPUBLAS, q3::execution::gpu::blas::HIPGPUBLAS, and q3::execution::gpu::blas::SYCLGPUBLAS.

Referenced by q3::GPUPlanewaveSubspaceManager::build_correction_vectors(), q3::GPUPlanewaveSubspaceManager::collapse_subspace(), q3::GPUPlanewaveSubspaceManager::extract_eigenvectors(), q3::GPUPlanewaveSubspaceManager::orthogonalize_batch(), and q3::GPUPlanewaveSubspaceManager::update_subspace_matrices().

◆ gemm_single()

virtual void q3::execution::gpu::blas::GPUBLAS::gemm_single	(	char	transpose_a,
		char	transpose_b,
		int	m,
		int	n,
		int	k,
		ComplexSingle	alpha,
		const ComplexSingle *	matrix_a,
		int	leading_dimension_a,
		const ComplexSingle *	matrix_b,
		int	leading_dimension_b,
		ComplexSingle	beta,
		ComplexSingle *	matrix_c,
		int	leading_dimension_c
	)		const

pure virtual

Complex single-precision general matrix-matrix multiply (CGEMM). matrix_c := alpha * op(matrix_a) * op(matrix_b) + beta * matrix_c

Implemented in q3::execution::gpu::blas::CudaGPUBLAS, q3::execution::gpu::blas::HIPGPUBLAS, and q3::execution::gpu::blas::SYCLGPUBLAS.

◆ norm()

virtual double q3::execution::gpu::blas::GPUBLAS::norm	(	int	number_of_elements,
		const Complex *	vector_x
	)		const

pure virtual

Euclidean norm of a complex double-precision vector.

Synchronizes before returning the scalar result.

Parameters

number_of_elements	Length of the vector.
vector_x	Input vector.

Returns: The real-valued norm.

Implemented in q3::execution::gpu::blas::CudaGPUBLAS, q3::execution::gpu::blas::HIPGPUBLAS, and q3::execution::gpu::blas::SYCLGPUBLAS.

Referenced by q3::execution::PlanewaveGPUExecutionPolicy::norm().

◆ norm_single()

virtual float q3::execution::gpu::blas::GPUBLAS::norm_single	(	int	number_of_elements,
		const ComplexSingle *	vector_x
	)		const

pure virtual

Euclidean norm of a complex single-precision vector.

Synchronizes before returning the scalar result.

Parameters

number_of_elements	Length of the vector.
vector_x	Input vector.

Returns: The real-valued norm (float).

Implemented in q3::execution::gpu::blas::CudaGPUBLAS, q3::execution::gpu::blas::HIPGPUBLAS, and q3::execution::gpu::blas::SYCLGPUBLAS.

◆ scale()

virtual void q3::execution::gpu::blas::GPUBLAS::scale	(	int	number_of_elements,
		Complex	scaling_factor,
		Complex *	vector_x
	)		const

pure virtual

Scale a complex double-precision vector: vector_x := scaling_factor * vector_x.

Parameters

number_of_elements	Length of the vector.
scaling_factor	Complex scaling coefficient.
vector_x	Vector to scale (in-place).

Implemented in q3::execution::gpu::blas::CudaGPUBLAS, q3::execution::gpu::blas::HIPGPUBLAS, and q3::execution::gpu::blas::SYCLGPUBLAS.

Referenced by q3::GPUPlanewaveSubspaceManager::build_correction_vectors().

Member Data Documentation

◆ device_backend_

std::shared_ptr<device::GPUDeviceBackend> q3::execution::gpu::blas::GPUBLAS::device_backend_

protected

Definition at line 212 of file init.hpp.

Referenced by q3::execution::gpu::blas::SYCLGPUBLAS::SYCLGPUBLAS(), device_backend(), and device_backend().

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

src/layer_0_core/math/linear_algebra/gpu/blas/init.hpp

Public Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ GPUBLAS()

◆ ~GPUBLAS()

Member Function Documentation

◆ axpy()

◆ device_backend() [1/2]

◆ device_backend() [2/2]

◆ dgemm()

◆ dot_product()

◆ dot_product_single()

◆ gemm()

◆ gemm_single()

◆ norm()

◆ norm_single()

◆ scale()

Member Data Documentation

◆ device_backend_