The CUDA IPC (interprocess communication) memory object, which internally contains data pointers to CUDA IPC memory. It is be useful for efficient all-reduce implementation. More...

#include <cuda_ipc_memory.h>

Inheritance diagram for tvm::runtime::cuda_ipc::CUDAIPCMemoryObj:

Collaboration diagram for tvm::runtime::cuda_ipc::CUDAIPCMemoryObj:

Public Member Functions
	TVM_DECLARE_BASE_OBJECT_INFO (CUDAIPCMemoryObj, Object)

Public Attributes
int	num_workers
	The number of GPU workers. More...

int	worker_id
	The worker id corresponding to this IPC memory object. More...

std::vector< void * >	remote_data
	The data pointers of all all-reduce inputs. It has "num_workers" pointers. The i-th pointer is the data pointer on worker i. If "i != worker_id", the pointer is an IPC data pointer. Otherwise, the pointer is a local CUDA data pointer. More...

std::vector< void * >	barrier_in
	The pointers to input barrier signals of all workers for all-reduce. It has "num_workers" pointers, and the pointer arrangement is the same as "remote_data". More...

std::vector< void * >	barrier_out
	The pointers to output barrier signals of all workers for all-reduce. It has "num_workers" pointers, and the pointer arrangement is the same as "remote_data". More...

int	barrier_flag
	The integer buffer flag for all-reduce. More...

Static Public Attributes
static constexpr const char *	_type_key = "tvm.runtime.disco.cuda_ipc_memory"

Detailed Description

The CUDA IPC (interprocess communication) memory object, which internally contains data pointers to CUDA IPC memory. It is be useful for efficient all-reduce implementation.

Note: Right now the class members are closely tied with customized all-reduce kernel. They may also be extended for other uses in the future.

Member Function Documentation

◆ TVM_DECLARE_BASE_OBJECT_INFO()

tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::TVM_DECLARE_BASE_OBJECT_INFO	(	CUDAIPCMemoryObj	,
		Object
	)

Member Data Documentation

◆ _type_key

constexpr const char* tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::_type_key = "tvm.runtime.disco.cuda_ipc_memory"

staticconstexpr

◆ barrier_flag

int tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::barrier_flag

The integer buffer flag for all-reduce.

◆ barrier_in

std::vector<void*> tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::barrier_in

The pointers to input barrier signals of all workers for all-reduce. It has "num_workers" pointers, and the pointer arrangement is the same as "remote_data".

◆ barrier_out

std::vector<void*> tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::barrier_out

The pointers to output barrier signals of all workers for all-reduce. It has "num_workers" pointers, and the pointer arrangement is the same as "remote_data".

◆ num_workers

int tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::num_workers

The number of GPU workers.

◆ remote_data

std::vector<void*> tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::remote_data

The data pointers of all all-reduce inputs. It has "num_workers" pointers. The i-th pointer is the data pointer on worker i. If "i != worker_id", the pointer is an IPC data pointer. Otherwise, the pointer is a local CUDA data pointer.

◆ worker_id

int tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::worker_id

The worker id corresponding to this IPC memory object.

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

include/tvm/runtime/disco/cuda_ipc_memory.h

Public Member Functions

Public Attributes

Static Public Attributes