triton_cpp::CudaSharedMemoryRegion Class Reference

Own a CUDA device allocation exportable through a CUDA IPC handle. More...

#include <cuda_shm.hpp>

Public Member Functions
	CudaSharedMemoryRegion (const std::string &name, std::int64_t size)
	Allocate a CUDA shared-memory region on the current device.
	~CudaSharedMemoryRegion ()
	Release the owned CUDA device allocation.
	CudaSharedMemoryRegion (const CudaSharedMemoryRegion &)=delete
CudaSharedMemoryRegion &	operator= (const CudaSharedMemoryRegion &)=delete
	CudaSharedMemoryRegion (CudaSharedMemoryRegion &&)=delete
CudaSharedMemoryRegion &	operator= (CudaSharedMemoryRegion &&)=delete
uint8_t *	getDeviceAddress () const
const cudaIpcMemHandle_t &	getIpcHandle () const
std::size_t	getDeviceId () const
std::int64_t	getSize () const
const std::string &	getName () const

Detailed Description

Own a CUDA device allocation exportable through a CUDA IPC handle.

The allocation is created on the current CUDA device and released when the region is destroyed. Regions are deliberately neither copyable nor movable.

Definition at line 61 of file cuda_shm.hpp.

Constructor & Destructor Documentation

CudaSharedMemoryRegion() [1/3]

triton_cpp::CudaSharedMemoryRegion::CudaSharedMemoryRegion ( const std::string & name, std::int64_t size )

inline

Allocate a CUDA shared-memory region on the current device.

Parameters

name	Name used when registering the region with Triton.
size	Allocation size in bytes.

Exceptions

std::runtime_error

if allocation or IPC handle creation fails.

Definition at line 69 of file cuda_shm.hpp.

                                                                 : name_{name}, size_{size} {
    throw_on_cuda_error(cudaGetDevice(&device_id_), "cudaGetDevice");
    throw_on_cuda_error(cudaMalloc(reinterpret_cast<void**>(&device_ptr_), static_cast<std::size_t>(size_)), "cudaMalloc");
    throw_on_cuda_error(cudaIpcGetMemHandle(&ipc_handle_, device_ptr_), "cudaIpcGetMemHandle");
  }

References triton_cpp::throw_on_cuda_error().

~CudaSharedMemoryRegion()

triton_cpp::CudaSharedMemoryRegion::~CudaSharedMemoryRegion ( )

inline

Release the owned CUDA device allocation.

Definition at line 76 of file cuda_shm.hpp.

                            {
    if (device_ptr_ != nullptr) {
      cudaFree(device_ptr_);
    }
  }

CudaSharedMemoryRegion() [2/3]

triton_cpp::CudaSharedMemoryRegion::CudaSharedMemoryRegion ( const CudaSharedMemoryRegion & )

delete

CudaSharedMemoryRegion() [3/3]

triton_cpp::CudaSharedMemoryRegion::CudaSharedMemoryRegion ( CudaSharedMemoryRegion && )

delete

Member Function Documentation

getDeviceAddress()

uint8_t * triton_cpp::CudaSharedMemoryRegion::getDeviceAddress ( ) const

inline

Returns

Device address of the allocation.

Definition at line 88 of file cuda_shm.hpp.

{ return device_ptr_; }

getDeviceId()

std::size_t triton_cpp::CudaSharedMemoryRegion::getDeviceId ( ) const

inline

Returns

CUDA device ordinal that owns the allocation.

Definition at line 92 of file cuda_shm.hpp.

{ return static_cast<std::size_t>(device_id_); }

getIpcHandle()

const cudaIpcMemHandle_t & triton_cpp::CudaSharedMemoryRegion::getIpcHandle ( ) const

inline

Returns

CUDA IPC handle for importing the allocation.

Definition at line 90 of file cuda_shm.hpp.

{ return ipc_handle_; }

getName()

const std::string & triton_cpp::CudaSharedMemoryRegion::getName ( ) const

inline

Returns

Name used to register this region with Triton.

Definition at line 96 of file cuda_shm.hpp.

{ return name_; }

getSize()

std::int64_t triton_cpp::CudaSharedMemoryRegion::getSize ( ) const

inline

Returns

Allocation size in bytes.

Definition at line 94 of file cuda_shm.hpp.

{ return size_; }

operator=() [1/2]

CudaSharedMemoryRegion & triton_cpp::CudaSharedMemoryRegion::operator= ( const CudaSharedMemoryRegion & )

delete

operator=() [2/2]

CudaSharedMemoryRegion & triton_cpp::CudaSharedMemoryRegion::operator= ( CudaSharedMemoryRegion && )

delete

---

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

• include/triton_cpp/cuda_shm.hpp