hmlp::Worker Class Reference

#include <thread.hpp>

Public Member Functions
	Worker ()
	Device implementation. More...
	Worker (thread_communicator *my_comm)
void	Communicator (thread_communicator *comm)
bool	Master ()
void	Barrier ()
void	InitWithCommunicator (thread_communicator *comm, size_t tid, size_t gid)
Worker	Split ()
template<typename Arg >
void	Bcast (Arg &buffer)
template<int ALIGN_SIZE, typename T >
T *	AllocateSharedMemory (size_t count)
template<typename T >
void	FreeSharedMemory (T *ptr)
size_t	BalanceOver1DGangs (size_t n, size_t default_size, size_t nb)
tuple< size_t, size_t, size_t >	DistributeOver1DGangs (size_t beg, size_t end, size_t nb)
tuple< size_t, size_t, size_t >	DistributeOver1DThreads (size_t beg, size_t end, size_t nb)
void	SetDevice (class Device *device)
class Device *	GetDevice ()
bool	Execute (class Task *task)
	The work executes the task in the runtime system. I left some code commented out because there is no GPU support now. With GPUs (or some distributed devices), we need to first gather data before the execution. More...
void	WaitExecute ()
	Pose a barrier if the device owned by this worker is performing asybchronous execution. More...
float	EstimateCost (class Task *task)

Public Attributes
int	tid = 0
int	gid = 0
int	child_gid = 0
int	jc_id
int	pc_id
int	ic_id
int	jr_id
int	ic_jr
int	jc_nt
int	pc_nt
int	ic_nt
int	jr_nt
thread_communicator *	my_comm
thread_communicator *	jc_comm
thread_communicator *	pc_comm
thread_communicator *	ic_comm
thread_communicator *	comm = NULL
class Scheduler *	scheduler

Detailed Description

end class thread_communicator

Constructor & Destructor Documentation

hmlp::Worker::Worker

(

)

Device implementation.

Worker implementation

Member Function Documentation

bool hmlp::Worker::Execute

(

class Task *

task

)

The work executes the task in the runtime system. I left some code commented out because there is no GPU support now. With GPUs (or some distributed devices), we need to first gather data before the execution.

Loop over each task in the batch. Notice that this may be an asynchronous execution. Wait for all tasks in the batch to terminate. Loop over each task in the batch.

Parameters

*task

The current task pointer.

Some tasks may be in "EXECUTED" status.

Move to the next task in the batch

Wait for all tasks in the batch to terminate.

Move to the next task in the batch

Set my current executing task to NULL.

class Device * hmlp::Worker::GetDevice

(

)

Return the device pointer attached to the worker.

void hmlp::Worker::SetDevice

(

class Device *

device

)

Assign a device to this worker.

Worker hmlp::Worker::Split

(

)

By default, we split threads evenly usine "close" affinity. Threads with the same color will be in the same subcomm.

example: (n_splits=2)

tid 0 1 2 3 4 5 6 7 color 0 0 0 0 1 1 1 1 (gang id) first 0 0 0 0 4 4 4 4 last 4 4 4 4 8 8 8 8 child_tid 0 1 2 3 0 1 2 3 4 4 4 4 4 4 4 4 (child_n_threads)

make sure all threads have the new communicator

void hmlp::Worker::WaitExecute

(

)

Pose a barrier if the device owned by this worker is performing asybchronous execution.

end Worker::Execute()

---

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

• /home/chenhany/Projects/hmlp/frame/base/thread.hpp
• /home/chenhany/Projects/hmlp/frame/base/thread.cpp