DeviceMesh Class Reference

This class wraps a vector of per-device IDs and CUDA streams and holds inter-device connectivity information and NCCL comms. More...

#include <nanovdb/cuda/DeviceMesh.h>

Public Member Functions
	DeviceMesh ()
	Constructs a device mesh for all devices on the host and initializes a CUDA stream (and NCCL communicator) for each device.
	~DeviceMesh ()
	Destroys the per-device CUDA stream and finalizes and destroys the per-device NCCL communicators.
	DeviceMesh (const DeviceMesh &)=delete
	Disallow copy-construction.
	DeviceMesh (DeviceMesh &&) noexcept
	Move constructor. Underlying CUDA streams and NCCL communicators are not reinitialized.
DeviceMesh &	operator= (const DeviceMesh &)=delete
	Disallow copy-assignment.
DeviceMesh &	operator= (DeviceMesh &&) noexcept
	Move assignment. Underlying CUDA streams and NCCL communicators are not reinitialized.
size_type	deviceCount () const
	Returns the number of devices.
const DeviceNode &	operator[] (int deviceId) const
	Returns a const reference to the DeviceNode for a particular device.
const_iterator	begin () const
	Returns an iterator to the first device node.
const_iterator	end () const
	Returns an iterator past the last device node.
bool	canAccessPeer (int deviceId, int peerId) const
	Returns whether or not peer to peer access is supported between two devices.

Detailed Description

This class wraps a vector of per-device IDs and CUDA streams and holds inter-device connectivity information and NCCL comms.

Constructor & Destructor Documentation

DeviceMesh() [1/3]

DeviceMesh ( )

inline

Constructs a device mesh for all devices on the host and initializes a CUDA stream (and NCCL communicator) for each device.

~DeviceMesh()

~DeviceMesh ( )

inline

Destroys the per-device CUDA stream and finalizes and destroys the per-device NCCL communicators.

DeviceMesh() [2/3]

DeviceMesh ( const DeviceMesh & )

delete

Disallow copy-construction.

DeviceMesh() [3/3]

DeviceMesh ( DeviceMesh && other )

inlinenoexcept

Move constructor. Underlying CUDA streams and NCCL communicators are not reinitialized.

Parameters

other

DeviceMesh instance that will be moved into this DeviceMesh.

Member Function Documentation

begin()

const_iterator begin ( ) const

inline

Returns an iterator to the first device node.

canAccessPeer()

bool canAccessPeer ( int deviceId, int peerId ) const

inline

Returns whether or not peer to peer access is supported between two devices.

deviceCount()

size_type deviceCount ( ) const

inline

Returns the number of devices.

end()

const_iterator end ( ) const

inline

Returns an iterator past the last device node.

operator=() [1/2]

DeviceMesh & operator= ( const DeviceMesh & )

delete

Disallow copy-assignment.

operator=() [2/2]

DeviceMesh & operator= ( DeviceMesh && other )

inlinenoexcept

Move assignment. Underlying CUDA streams and NCCL communicators are not reinitialized.

Parameters

other

DeviceMesh instance that will be moved into this DeviceMesh.

operator[]()

const DeviceNode & operator[] ( int deviceId ) const

inline

Returns a const reference to the DeviceNode for a particular device.