PIMPL class for CUDA or HIP stream. More...

#include <Stream.hh>

Classes
struct	Impl
	PIMPL class. More...

Public Types
Type aliases
using	StreamT = std::nullptr_t
	Stream implementation is unavailable.

using	ResourceT = std::nullptr_t
	Stream implementation is unavailable.

using	HostKernel = void()(void )
	Stream implementation is unavailable.

Public Member Functions
	Stream ()
	Construct by creating a stream with the active device context.

	Stream (std::nullptr_t)
	Construct a null stream.

	Stream (Device const &device)
	Construct a stream for the given device.

	Stream (Stream const &)=delete

Stream &	operator= (Stream const &)=delete

	Stream (Stream &&)=default

Stream &	operator= (Stream &&)=default

	operator bool () const
	Whether the stream is valid (not null or moved-from).

StreamT	get () const
	Get the CUDA stream pointer.

ResourceT &	memory_resource ()
	Get the Thrust async allocation resource.

void *	malloc_async (std::size_t bytes)
	Allocate memory asynchronously on this stream if possible.

void	free_async (void *ptr)
	Free memory asynchronously on this stream if possible.

void	sync () const
	Block host execution until stream operations are all complete.

void	wait (DeviceEvent const &d)
	Block stream execution until the event completes.

void	launch_host_func (HostKernel func, void *data)
	Enqueue delayed execution of a host function.

Detailed Description

PIMPL class for CUDA or HIP stream.

This creates/destroys a stream on construction/destruction and provides accessors to low-level stream-related functionality. This class will typically be accessed only by low-level device implementations or advanced kernels that need to interact with the device stream.

States

Constructed: A valid stream created with an active device. The stream can be used for device operations.
Null: Explicitly constructed with nullptr. No stream is created, but the object is in a valid null state. Operations are no-ops.

Warning: This class interface changes based on available headers. Because the CUDA/HIP stream type are only defined when those paths are included and available (which isn't true for all Celeritas code) we hide the stream unless DeviceRuntimeApi has been included or the file is being compiled by a CUDA/HIP compiler. Worse, our stream class has to manage a Thrust memory resource, and in newer versions of rocthrust, its headers cannot be included at all by a non-HIP compiler. Because this class forward declares the memory resource, downstream uses must include corecel/sys/detail/AsyncMemoryResource.device.hh .

Constructor & Destructor Documentation

◆ Stream() [1/2]

celeritas::Stream::Stream ( )

Construct by creating a stream with the active device context.

Precondition: A device must be active and configured.

Deprecated:: This constructor is ambiguous: remove it.

◆ Stream() [2/2]

celeritas::Stream::Stream ( Device const & device )

explicit

Construct a stream for the given device.

This delegates to the default constructor, which creates a stream on the active device. The device reference is used to enforce that a valid device has been activated before stream creation.

Precondition: The device must be valid and active.

Member Function Documentation

◆ malloc_async()

void * celeritas::Stream::malloc_async ( std::size_t bytes )

Allocate memory asynchronously on this stream if possible.

HIP 5.1 and lower does not support async allocation.

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

Classes