Mila::Dnn::Compute::CudaDevice Class Reference

Class representing a CUDA compute device. More...

Inheritance diagram for Mila::Dnn::Compute::CudaDevice:

Collaboration diagram for Mila::Dnn::Compute::CudaDevice:

Public Types
using	MANAGED_MR = CudaManagedMemoryResource
	Type alias for managed memory resource.
using	MR = CudaMemoryResource
	Type alias for device memory resource.
using	PINNED_MR = CudaPinnedMemoryResource
	Type alias for pinned memory resource.

Public Member Functions
	CudaDevice (int device_id)
	Constructs a CudaDevice with specified device ID.
int	getDeviceId () const
	Gets the CUDA device ID.
constexpr DeviceType	getDeviceType () const override
	Gets the type of this compute device.
std::string	getName () const override
	Gets the name of this CUDA device.
const DeviceProps &	getProperties () const
	Gets the properties of this CUDA device.
bool	isFp16Supported ()
	Checks if the device supports FP16 precision.
bool	isFp4Supported ()
	Checks if the device supports FP4 precision.
bool	isFp8Supported ()
	Checks if the device supports FP8 precision.
Public Member Functions inherited from Mila::Dnn::Compute::ComputeDevice
virtual	~ComputeDevice ()=default
	Virtual destructor.

Static Public Member Functions
static void	registerDevices ()
	Registers all available CUDA devices with the DeviceRegistry.

Private Attributes
int	device_id_
	The CUDA device ID.
DeviceProps	props_
	Device properties.

Static Private Attributes
static bool	registered_ { false }
	Flag indicating if devices have been registered.

Detailed Description

Class representing a CUDA compute device.

This class provides an interface to interact with NVIDIA CUDA-capable GPUs within the Mila framework. It handles device properties retrieval and registration in the DeviceRegistry. Device activation is managed by the DeviceContext class.

Member Typedef Documentation

MANAGED_MR

using Mila::Dnn::Compute::CudaDevice::MANAGED_MR = CudaManagedMemoryResource

Type alias for managed memory resource.

MR

using Mila::Dnn::Compute::CudaDevice::MR = CudaMemoryResource

Type alias for device memory resource.

PINNED_MR

using Mila::Dnn::Compute::CudaDevice::PINNED_MR = CudaPinnedMemoryResource

Type alias for pinned memory resource.

Constructor & Destructor Documentation

CudaDevice()

Mila::Dnn::Compute::CudaDevice::CudaDevice ( int  device_id )

inlineexplicit

Constructs a CudaDevice with specified device ID.

Parameters

device_id

The CUDA device ID to initialize.

Member Function Documentation

getDeviceId()

int Mila::Dnn::Compute::CudaDevice::getDeviceId ( ) const

inline

Gets the CUDA device ID.

Returns

int The device ID for this CUDA device.

getDeviceType()

constexpr DeviceType Mila::Dnn::Compute::CudaDevice::getDeviceType ( ) const

inlineconstexproverridevirtual

Gets the type of this compute device.

Returns

DeviceType The device type (CUDA).

Implements Mila::Dnn::Compute::ComputeDevice.

getName()

std::string Mila::Dnn::Compute::CudaDevice::getName ( ) const

inlineoverridevirtual

Gets the name of this CUDA device.

Returns

std::string The device name in format "CUDA:<device_id>".

Implements Mila::Dnn::Compute::ComputeDevice.

getProperties()

const DeviceProps & Mila::Dnn::Compute::CudaDevice::getProperties ( ) const

inline

Gets the properties of this CUDA device.

Returns

const DeviceProps& Reference to the device properties.

isFp16Supported()

bool Mila::Dnn::Compute::CudaDevice::isFp16Supported ( )

inline

Checks if the device supports FP16 precision.

Returns

bool True if FP16 is supported, false otherwise.

isFp4Supported()

bool Mila::Dnn::Compute::CudaDevice::isFp4Supported ( )

inline

Checks if the device supports FP4 precision.

Returns

bool True if FP4 is supported, false otherwise.

isFp8Supported()

bool Mila::Dnn::Compute::CudaDevice::isFp8Supported ( )

inline

Checks if the device supports FP8 precision.

Returns

bool True if FP8 is supported, false otherwise.

registerDevices()

static void Mila::Dnn::Compute::CudaDevice::registerDevices ( )

inlinestatic

Registers all available CUDA devices with the DeviceRegistry.

This method discovers all CUDA devices in the system and registers them with the DeviceRegistry for later instantiation. It performs the registration only once, even if called multiple times.

Here is the call graph for this function:

Member Data Documentation

device_id_

int Mila::Dnn::Compute::CudaDevice::device_id_

private

The CUDA device ID.

props_

DeviceProps Mila::Dnn::Compute::CudaDevice::props_

private

Device properties.

registered_

bool Mila::Dnn::Compute::CudaDevice::registered_ { false }

inlinestaticprivate

Flag indicating if devices have been registered.

---

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

• /home/runner/work/Mila/Mila/Mila/Src/Dnn/Compute/Devices/Cuda/CudaDevice.ixx