uni20::cuda Namespace Reference

Classes
class	AccessCompletion
	One operation-tail completion shared by multiple CUDA buffer accesses. More...
class	BlockingReadAccess
	Scoped read-only CUDA pointer access after a host-side synchronization. More...
class	BlockingWriteAccess
	Scoped read/write CUDA pointer access after a host-side synchronization. More...
class	Completion
	Shared completion token for one submitted CUDA operation. More...
class	CudaBuffer
	Move-only logical CUDA buffer and completion state. More...
struct	CudaBufferRange
	One logical range used to partition a CUDA allocation. More...
class	CudaBufferView
	Non-owning opaque view into one logical CUDA buffer. More...
class	CudaComplexReference
	Assignable CUDA execution proxy over one persistent complex value. More...
struct	CudaConjugate
	Named CUDA-callable complex conjugation operation. More...
struct	CudaConjugatingPointerAccessor
	Read-only CUDA accessor that conjugates persistent complex storage. More...
struct	CudaPointerAccessor
	Accessor for a CUDA mdspan whose device pointer has been leased. More...
struct	CudaPointerAccessor< uni20::detail::standard_complex< Real > >
	Mutable CUDA accessor for persistent uni20::complex storage.
struct	CudaPointerAccessor< uni20::detail::standard_complex< Real > const >
	Read-only CUDA accessor for persistent uni20::complex storage.
class	CudaRuntimeError
	Exception carrying a failed CUDA runtime operation and status. More...
class	Device
	Validated CUDA device identity with cached immutable capabilities. More...
struct	DeviceCapabilities
	Immutable hardware properties cached for one CUDA device ordinal. More...
class	DeviceResources
	Mutable stream and provider resources associated with one CUDA device. More...
class	OwningReadAccess
	Read access that owns a moved CUDA buffer. More...
class	PartitionedCudaBuffer
	One CUDA allocation partitioned into independent logical buffers. More...
class	PartitionedReadAccess
	Scoped read access to every child of one partitioned CUDA allocation. More...
class	PartitionedWriteAccess
	Scoped write access to every child of one partitioned CUDA allocation. More...
class	ReadAccess
	Scoped read-only access to a CUDA buffer on one stream. More...
class	ResourceLease
	Move-only exclusive lease of one provider-resource pool slot. More...
class	ResourcePool
	Fixed-capacity pool of preconstructed provider resources. More...
class	Runtime
	Scoped owner of the process-wide CUDA resource installation. More...
struct	RuntimeConfig
	Configuration for one scoped process-wide CUDA runtime installation. More...
class	ScopedDevice
	Temporarily select one CUDA device and restore the previous device on destruction. More...
class	Stream
	Reference-counted lease of one CUDA stream-pool slot. More...
class	StreamPool
	Device-local pool whose streams become available only after their queued work completes. More...
class	WriteAccess
	Scoped read/write access to a CUDA buffer on one stream. More...

Concepts
concept	BufferMdspec
	CUDA-accessible mdspec backed by a CudaBufferView descriptor.

Functions
void	check (cudaError_t status, std::string_view operation, std::optional< int > device=std::nullopt, std::source_location where=std::source_location::current())
	Raise when a CUDA runtime operation did not return cudaSuccess.
template<class ElementType > requires (!std::is_const_v<ElementType>)
UNI20_HOST_DEVICE constexpr auto	const_accessor (CudaPointerAccessor< ElementType > const &)
	Produce a read-only CUDA pointer accessor without a generic lvalue adaptor.
DeviceResources &	device_resources ()
	Return canonical resources for the installed runtime's default device.
DeviceResources &	device_resources (int device)
	Return canonical resources for an enrolled CUDA device.
presentation::report_builder	diagnostic_report (CudaRuntimeError const &error)
	Build a presentation report for a CUDA runtime failure.
Runtime	initialize (RuntimeConfig config={})
	Install and return a scoped process-wide CUDA runtime owner.
bool	is_initialized () noexcept
	Return whether a process-wide CUDA runtime is currently installed.
void	raise_runtime_error (cudaError_t status, std::string_view operation, std::optional< int > device=std::nullopt, std::source_location where=std::source_location::current())
	Raise a structured CUDA runtime failure through Uni20's error boundary.
Runtime &	runtime ()
	Return the installed runtime or fail when CUDA was not initialized.

Function Documentation

check()

void uni20::cuda::check	(	cudaError_t	status,
		std::string_view	operation,
		std::optional< int >	device,
		std::source_location	where
	)

Raise when a CUDA runtime operation did not return cudaSuccess.

const_accessor()

template<class ElementType >
requires (!std::is_const_v<ElementType>)

UNI20_HOST_DEVICE constexpr auto uni20::cuda::const_accessor ( CudaPointerAccessor< ElementType > const &  )

constexpr

Produce a read-only CUDA pointer accessor without a generic lvalue adaptor.

device_resources() [1/2]

DeviceResources & uni20::cuda::device_resources

(

)

Return canonical resources for the installed runtime's default device.

device_resources() [2/2]

DeviceResources & uni20::cuda::device_resources

(

int

device

)

Return canonical resources for an enrolled CUDA device.

diagnostic_report()

presentation::report_builder uni20::cuda::diagnostic_report ( CudaRuntimeError const &  error )

inline

Build a presentation report for a CUDA runtime failure.

initialize()

Runtime uni20::cuda::initialize

(

RuntimeConfig

config = {}

)

Install and return a scoped process-wide CUDA runtime owner.

Only one installation may be active. Retain the returned object at the application startup boundary for as long as CUDA objects exist.

is_initialized()

bool uni20::cuda::is_initialized ( )

noexcept

Return whether a process-wide CUDA runtime is currently installed.

raise_runtime_error()

void uni20::cuda::raise_runtime_error	(	cudaError_t	status,
		std::string_view	operation,
		std::optional< int >	device,
		std::source_location	where
	)

Raise a structured CUDA runtime failure through Uni20's error boundary.

runtime()

Runtime & uni20::cuda::runtime

(

)

Return the installed runtime or fail when CUDA was not initialized.