Classes
class	CublasError
	Exception carrying a failed cuBLAS operation and device identity. More...

class	ExecutionAcquireAwaiter
	Awaiter that reserves a cuBLAS handle before waiting for an idle stream. More...

class	ExecutionLease
	Move-only dynamic pairing of one cuBLAS handle slot and one idle stream. More...

class	ExecutionPool
	Device-local cuBLAS handle pool combined with a shared idle-stream pool. More...

class	HandleSlot
	One expensive reusable cuBLAS handle and its handle-local state. More...

Concepts
concept	CublasScalar
	Scalar types supported by the cuBLAS GEMM wrappers.

concept	CublasLevelOneScalar
	Scalar types supported by the cuBLAS level-one wrappers.

Functions
ExecutionAcquireAwaiter	acquire_execution (ExecutionPool &pool) noexcept
	Acquire a cuBLAS handle and idle stream without blocking a CUDA task.

void	check (cublasStatus_t status, std::string_view operation, int device, std::source_location where=std::source_location::current())
	Raise unless a cuBLAS call returned `CUBLAS_STATUS_SUCCESS`.

presentation::report_builder	diagnostic_report (CublasError const &error)
	Build a presentation report for a cuBLAS provider failure.

template<CublasLevelOneScalar Scalar>
auto	dotc (ExecutionLease &execution, int size, Scalar const lhs, Scalar const rhs) -> Scalar
	Compute the conjugate-linear dot product and return a host scalar.

ExecutionPool &	execution_pool (cuda::DeviceResources &resources)
	Return the lazily constructed cuBLAS execution pool for one CUDA device.

ExecutionPool &	execution_pool (cuda::DeviceResources &resources, std::size_t handle_count)
	Configure or return the canonical cuBLAS execution pool for one device.

template<CublasScalar Scalar>
void	gemm (ExecutionLease &execution, char lhs_transform, char rhs_transform, int rows, int cols, int inner, Scalar alpha, Scalar const lhs, int lhs_leading_dimension, Scalar const rhs, int rhs_leading_dimension, Scalar beta, Scalar *output, int output_leading_dimension)
	Enqueue column-major GEMM through an exclusive handle/stream lease.

template<CublasLevelOneScalar Scalar>
auto	nrm2 (ExecutionLease &execution, int size, Scalar const *input) -> uni20::make_real_t< Scalar >
	Compute a Euclidean norm and return a host real scalar.

void	raise_error (cublasStatus_t status, std::string_view operation, int device, std::source_location where=std::source_location::current())
	Raise a structured cuBLAS provider failure.

std::string_view	status_name (cublasStatus_t status) noexcept
	Return a stable symbolic name for a cuBLAS status value.

Function Documentation

◆ acquire_execution()

ExecutionAcquireAwaiter uni20::cublas::acquire_execution ( ExecutionPool & pool )

inlinenoexcept

Acquire a cuBLAS handle and idle stream without blocking a CUDA task.

◆ check()

void uni20::cublas::check	(	cublasStatus_t	status,
		std::string_view	operation,
		int	device,
		std::source_location	where
	)

Raise unless a cuBLAS call returned CUBLAS_STATUS_SUCCESS.

◆ diagnostic_report()

presentation::report_builder uni20::cublas::diagnostic_report ( CublasError const & error )

inline

Build a presentation report for a cuBLAS provider failure.

◆ dotc()

template<CublasLevelOneScalar Scalar>

auto uni20::cublas::dotc	(	ExecutionLease &	execution,
		int	size,
		Scalar const *	lhs,
		Scalar const *	rhs
	)		-> Scalar

Compute the conjugate-linear dot product and return a host scalar.

The provider call uses host pointer mode. The operation stream is synchronized before the host result and any borrowed input storage may leave scope.

◆ execution_pool() [1/2]

ExecutionPool & uni20::cublas::execution_pool ( cuda::DeviceResources & resources )

Return the lazily constructed cuBLAS execution pool for one CUDA device.

The device resources own the pool and retain it until after every Tensor and operation using those resources has finished. Unless the pool was configured explicitly, it contains one handle per stream.

◆ execution_pool() [2/2]

ExecutionPool & uni20::cublas::execution_pool	(	cuda::DeviceResources &	resources,
		std::size_t	handle_count
	)

Configure or return the canonical cuBLAS execution pool for one device.

This first-use configuration must precede concurrent provider use. Repeated calls must request the same handle count. The handle count cannot exceed the device's stream capacity.

Parameters

resources	Device resources that own the canonical pool.
handle_count	Number of reusable cuBLAS handles.

◆ gemm()

template<CublasScalar Scalar>

void uni20::cublas::gemm	(	ExecutionLease &	execution,
		char	lhs_transform,
		char	rhs_transform,
		int	rows,
		int	cols,
		int	inner,
		Scalar	alpha,
		Scalar const *	lhs,
		int	lhs_leading_dimension,
		Scalar const *	rhs,
		int	rhs_leading_dimension,
		Scalar	beta,
		Scalar *	output,
		int	output_leading_dimension
	)

Enqueue column-major GEMM through an exclusive handle/stream lease.

◆ nrm2()

template<CublasLevelOneScalar Scalar>

auto uni20::cublas::nrm2	(	ExecutionLease &	execution,
		int	size,
		Scalar const *	input
	)		-> uni20::make_real_t<Scalar>

Compute a Euclidean norm and return a host real scalar.

The operation stream is synchronized before the result returns.

◆ raise_error()

void uni20::cublas::raise_error	(	cublasStatus_t	status,
		std::string_view	operation,
		int	device,
		std::source_location	where
	)

Raise a structured cuBLAS provider failure.

◆ status_name()

std::string_view uni20::cublas::status_name ( cublasStatus_t status )

noexcept

Return a stable symbolic name for a cuBLAS status value.

Classes

Concepts

Functions