#include <dispatcher.hh>

Inheritance diagram for GPUDispatcher:

Public Types
typedef GPUDispatcherParams	Params

Public Types inherited from SimObject
typedef SimObjectParams	Params

Public Member Functions
	GPUDispatcher (const Params *p)

	~GPUDispatcher ()

void	serialize (CheckpointOut &cp) const override
	Serialize an object. More...

void	unserialize (CheckpointIn &cp) override
	Unserialize an object. More...

void	regStats () override
	Callback to set stat parameters. More...

void	setCommandProcessor (GPUCommandProcessor *gpu_cmd_proc)

void	setShader (Shader *new_shader)

void	exec ()

bool	isReachingKernelEnd (Wavefront *wf)

void	updateInvCounter (int kern_id, int val=-1)
	update the counter of oustanding inv requests for the kernel kern_id: kernel id val: +1/-1, increment or decrement the counter (default: -1) More...

bool	updateWbCounter (int kern_id, int val=-1)
	update the counter of oustanding wb requests for the kernel kern_id: kernel id val: +1/-1, increment or decrement the counter (default: -1) More...

int	getOutstandingWbs (int kern_id)
	get kernel's outstanding cache writeback requests More...

void	notifyWgCompl (Wavefront *wf)
	When an end program instruction detects that the last WF in a WG has completed it will call this method on the dispatcher. More...

void	scheduleDispatch ()

void	dispatch (HSAQueueEntry *task)
	After all relevant HSA data structures have been traversed/extracted from memory by the CP, dispatch() is called on the dispatcher. More...

HSAQueueEntry *	hsaTask (int disp_id)

Public Member Functions inherited from SimObject
const Params *	params () const

	SimObject (const Params *_params)

virtual	~SimObject ()

virtual const std::string	name () const

virtual void	init ()
	init() is called after all C++ SimObjects have been created and all ports are connected. More...

virtual void	loadState (CheckpointIn &cp)
	loadState() is called on each SimObject when restoring from a checkpoint. More...

virtual void	initState ()
	initState() is called on each SimObject when not restoring from a checkpoint. More...

virtual void	regProbePoints ()
	Register probe points for this object. More...

virtual void	regProbeListeners ()
	Register probe listeners for this object. More...

ProbeManager *	getProbeManager ()
	Get the probe manager for this object. More...

virtual Port &	getPort (const std::string &if_name, PortID idx=InvalidPortID)
	Get a port with a given name and index. More...

virtual void	startup ()
	startup() is the final initialization call before simulation. More...

DrainState	drain () override
	Provide a default implementation of the drain interface for objects that don't need draining. More...

virtual void	memWriteback ()
	Write back dirty buffers to memory using functional writes. More...

virtual void	memInvalidate ()
	Invalidate the contents of memory buffers. More...

void	serialize (CheckpointOut &cp) const override
	Serialize an object. More...

void	unserialize (CheckpointIn &cp) override
	Unserialize an object. More...

Public Member Functions inherited from EventManager
EventQueue *	eventQueue () const

void	schedule (Event &event, Tick when)

void	deschedule (Event &event)

void	reschedule (Event &event, Tick when, bool always=false)

void	schedule (Event *event, Tick when)

void	deschedule (Event *event)

void	reschedule (Event *event, Tick when, bool always=false)

void	wakeupEventQueue (Tick when=(Tick) -1)
	This function is not needed by the usual gem5 event loop but may be necessary in derived EventQueues which host gem5 on other schedulers. More...

void	setCurTick (Tick newVal)

	EventManager (EventManager &em)
	Event manger manages events in the event queue. More...

	EventManager (EventManager *em)

	EventManager (EventQueue *eq)

Public Member Functions inherited from Serializable
	Serializable ()

virtual	~Serializable ()

void	serializeSection (CheckpointOut &cp, const char *name) const
	Serialize an object into a new section. More...

void	serializeSection (CheckpointOut &cp, const std::string &name) const

void	unserializeSection (CheckpointIn &cp, const char *name)
	Unserialize an a child object. More...

void	unserializeSection (CheckpointIn &cp, const std::string &name)

Public Member Functions inherited from Drainable
DrainState	drainState () const
	Return the current drain state of an object. More...

virtual void	notifyFork ()
	Notify a child process of a fork. More...

Public Member Functions inherited from Stats::Group
	Group (Group parent, const char name=nullptr)
	Construct a new statistics group. More...

virtual	~Group ()

virtual void	resetStats ()
	Callback to reset stats. More...

virtual void	preDumpStats ()
	Callback before stats are dumped. More...

void	addStat (Stats::Info *info)
	Register a stat with this group. More...

const std::map< std::string, Group * > &	getStatGroups () const
	Get all child groups associated with this object. More...

const std::vector< Info * > &	getStats () const
	Get all stats associated with this object. More...

void	addStatGroup (const char name, Group block)
	Add a stat block as a child of this block. More...

const Info *	resolveStat (std::string name) const
	Resolve a stat by its name within this group. More...

	Group ()=delete

	Group (const Group &)=delete

Group &	operator= (const Group &)=delete

Private Attributes
Shader *	shader

GPUCommandProcessor *	gpuCmdProc

EventFunctionWrapper	tickEvent

std::unordered_map< int, HSAQueueEntry * >	hsaQueueEntries

std::queue< int >	execIds

std::queue< int >	doneIds

bool	dispatchActive

Stats::Scalar	numKernelLaunched

Stats::Scalar	cyclesWaitingForDispatch

Additional Inherited Members
Static Public Member Functions inherited from SimObject
static void	serializeAll (CheckpointOut &cp)
	Serialize all SimObjects in the system. More...

static SimObject *	find (const char *name)
	Find the SimObject with the given name and return a pointer to it. More...

Static Public Member Functions inherited from Serializable
static const std::string &	currentSection ()
	Gets the fully-qualified name of the active section. More...

static void	serializeAll (const std::string &cpt_dir)
	Serializes all the SimObjects. More...

static void	unserializeGlobals (CheckpointIn &cp)

Protected Member Functions inherited from Drainable
	Drainable ()

virtual	~Drainable ()

virtual void	drainResume ()
	Resume execution after a successful drain. More...

void	signalDrainDone () const
	Signal that an object is drained. More...

Protected Attributes inherited from SimObject
const SimObjectParams *	_params
	Cached copy of the object parameters. More...

Protected Attributes inherited from EventManager
EventQueue *	eventq
	A pointer to this object's event queue. More...

Detailed Description

Definition at line 60 of file dispatcher.hh.

Member Typedef Documentation

◆ Params

typedef GPUDispatcherParams GPUDispatcher::Params

Definition at line 63 of file dispatcher.hh.

Constructor & Destructor Documentation

◆ GPUDispatcher()

GPUDispatcher::GPUDispatcher ( const Params * p )

Definition at line 47 of file dispatcher.cc.

References exec().

◆ ~GPUDispatcher()

GPUDispatcher::~GPUDispatcher ( )

Definition at line 56 of file dispatcher.cc.

Member Function Documentation

◆ dispatch()

void GPUDispatcher::dispatch ( HSAQueueEntry * task )

After all relevant HSA data structures have been traversed/extracted from memory by the CP, dispatch() is called on the dispatcher.

This will schedule a dispatch event that, when triggered, will attempt to dispatch the WGs associated with the given task to the CUs.

Definition at line 127 of file dispatcher.cc.

References Clocked::clockPeriod(), curTick(), dispatchActive, HSAQueueEntry::dispatchId(), DPRINTF, execIds, hsaQueueEntries, HSAQueueEntry::kernelName(), numKernelLaunched, EventManager::schedule(), Event::scheduled(), shader, and tickEvent.

Referenced by GPUCommandProcessor::dispatchPkt().

◆ exec()

void GPUDispatcher::exec ( )

There are potentially multiple outstanding kernel launches. It is possible that the workgroups in a different kernel can fit on the GPU even if another kernel's workgroups cannot

dispatch work cannot start until the kernel's invalidate is completely finished; hence, kernel will always initiates invalidate first and keeps waiting until inv done

invalidate is still ongoing, put the kernel on the queue to retry later

if we failed try the next kernel, it may have smaller workgroups. put it on the queue to rety latter

Definition at line 144 of file dispatcher.cc.

References curTick(), cyclesWaitingForDispatch, Shader::dispatchWorkgroups(), doneIds, DPRINTF, execIds, hsaQueueEntries, Shader::impl_kern_launch_acq, Shader::prepareInvalidate(), shader, and Shader::updateContext().

Referenced by GPUDispatcher().

◆ getOutstandingWbs()

int GPUDispatcher::getOutstandingWbs ( int kern_id )

get kernel's outstanding cache writeback requests

Definition at line 286 of file dispatcher.cc.

References hsaQueueEntries.

Referenced by Shader::prepareFlush().

◆ hsaTask()

HSAQueueEntry * GPUDispatcher::hsaTask ( int disp_id )

Definition at line 76 of file dispatcher.cc.

References hsaQueueEntries.

◆ isReachingKernelEnd()

bool GPUDispatcher::isReachingKernelEnd ( Wavefront * wf )

whether the next workgroup is the final one in the kernel, +1 as we check first before taking action

Definition at line 232 of file dispatcher.cc.

References hsaQueueEntries, and Wavefront::kernId.

Referenced by Gcn3ISA::Inst_SOPP__S_ENDPGM::execute().

◆ notifyWgCompl()

void GPUDispatcher::notifyWgCompl ( Wavefront * wf )

When an end program instruction detects that the last WF in a WG has completed it will call this method on the dispatcher.

If we detect that this is the last WG for the given task, then we ring the completion signal, which is used by the CPU to synchronize with the GPU. The HSAPP is also notified that the task has completed so it can be removed from its task queues.

HACK: The semantics of the HSA signal is to decrement the current signal value. We cheat here and read out he value from main memory using functional access and then just DMA the decremented value. This is because the DMA controller does not currently support GPU atomics.

Definition at line 301 of file dispatcher.cc.

References Clocked::clockPeriod(), Wavefront::computeUnit, BaseBufferArg::copyIn(), ComputeUnit::cu_id, curTick(), HSADevice::dmaWriteVirt(), DPRINTF, HSAPacketProcessor::finishPkt(), gpuCmdProc, HSADevice::hsaPacketProc(), hsaQueueEntries, Wavefront::kernId, EventManager::schedule(), Event::scheduled(), shader, GPUCommandProcessor::system(), System::threads, tickEvent, and Wavefront::wgId.

Referenced by Gcn3ISA::Inst_SOPP__S_ENDPGM::execute(), and ComputeUnit::DataPort::recvTimingResp().

◆ regStats()

void GPUDispatcher::regStats ( )

overridevirtual

Callback to set stat parameters.

This callback is typically used for complex stats (e.g., distributions) that need parameters in addition to a name and a description. Stat names and descriptions should typically be set from the constructor usingo from the constructor using the ADD_STAT macro.

Reimplemented from Stats::Group.

Definition at line 61 of file dispatcher.cc.

References cyclesWaitingForDispatch, Stats::DataWrap< Derived, InfoProxyType >::desc(), SimObject::name(), Stats::DataWrap< Derived, InfoProxyType >::name(), and numKernelLaunched.

◆ scheduleDispatch()

void GPUDispatcher::scheduleDispatch ( )

Definition at line 357 of file dispatcher.cc.

References Clocked::clockPeriod(), curTick(), EventManager::schedule(), Event::scheduled(), shader, and tickEvent.

Referenced by Gcn3ISA::Inst_SOPP__S_ENDPGM::execute().

◆ serialize()

void GPUDispatcher::serialize ( CheckpointOut & cp ) const

overridevirtual

Serialize an object.

Output an object's state into the current checkpoint section.

Parameters

cp	Checkpoint state

Implements Serializable.

Definition at line 95 of file dispatcher.cc.

References Event::scheduled(), SERIALIZE_SCALAR, tickEvent, and Event::when().

◆ setCommandProcessor()

void GPUDispatcher::setCommandProcessor ( GPUCommandProcessor * gpu_cmd_proc )

Definition at line 83 of file dispatcher.cc.

References gpuCmdProc.

Referenced by GPUCommandProcessor::GPUCommandProcessor().

◆ setShader()

void GPUDispatcher::setShader ( Shader * new_shader )

Definition at line 89 of file dispatcher.cc.

References shader.

◆ unserialize()

void GPUDispatcher::unserialize ( CheckpointIn & cp )

overridevirtual

Unserialize an object.

Read an object's state from the current checkpoint section.

Parameters

cp	Checkpoint state

Implements Serializable.

Definition at line 106 of file dispatcher.cc.

References EventManager::deschedule(), EventManager::schedule(), Event::scheduled(), tickEvent, and UNSERIALIZE_SCALAR.

◆ updateInvCounter()

void GPUDispatcher::updateInvCounter	(	int	kern_id,
		int	val = `-1`
	)

update the counter of oustanding inv requests for the kernel kern_id: kernel id val: +1/-1, increment or decrement the counter (default: -1)

Definition at line 252 of file dispatcher.cc.

References Clocked::clockPeriod(), curTick(), hsaQueueEntries, EventManager::schedule(), Event::scheduled(), shader, tickEvent, and X86ISA::val.

Referenced by Shader::prepareInvalidate(), and ComputeUnit::DataPort::recvTimingResp().

◆ updateWbCounter()

bool GPUDispatcher::updateWbCounter	(	int	kern_id,
		int	val = `-1`
	)

update the counter of oustanding wb requests for the kernel kern_id: kernel id val: +1/-1, increment or decrement the counter (default: -1)

return true if all wbs are done for the kernel

Definition at line 272 of file dispatcher.cc.

References hsaQueueEntries, and X86ISA::val.

Referenced by Shader::prepareFlush(), and ComputeUnit::DataPort::recvTimingResp().

Member Data Documentation

◆ cyclesWaitingForDispatch

Stats::Scalar GPUDispatcher::cyclesWaitingForDispatch

private

Definition at line 96 of file dispatcher.hh.

Referenced by exec(), and regStats().

◆ dispatchActive

bool GPUDispatcher::dispatchActive

private

Definition at line 93 of file dispatcher.hh.

Referenced by dispatch().

◆ doneIds

std::queue<int> GPUDispatcher::doneIds

private

Definition at line 91 of file dispatcher.hh.

Referenced by exec().

◆ execIds

std::queue<int> GPUDispatcher::execIds

private

Definition at line 89 of file dispatcher.hh.

Referenced by dispatch(), and exec().

◆ gpuCmdProc

GPUCommandProcessor* GPUDispatcher::gpuCmdProc

private

Definition at line 85 of file dispatcher.hh.

Referenced by notifyWgCompl(), and setCommandProcessor().

◆ hsaQueueEntries

std::unordered_map<int, HSAQueueEntry*> GPUDispatcher::hsaQueueEntries

private

Definition at line 87 of file dispatcher.hh.

Referenced by dispatch(), exec(), getOutstandingWbs(), hsaTask(), isReachingKernelEnd(), notifyWgCompl(), updateInvCounter(), and updateWbCounter().

◆ numKernelLaunched

Stats::Scalar GPUDispatcher::numKernelLaunched

private

Definition at line 95 of file dispatcher.hh.

Referenced by dispatch(), and regStats().

◆ shader

Shader* GPUDispatcher::shader

private

Definition at line 84 of file dispatcher.hh.

Referenced by dispatch(), exec(), notifyWgCompl(), scheduleDispatch(), setShader(), and updateInvCounter().

◆ tickEvent

EventFunctionWrapper GPUDispatcher::tickEvent

private

Definition at line 86 of file dispatcher.hh.

Referenced by dispatch(), notifyWgCompl(), scheduleDispatch(), serialize(), unserialize(), and updateInvCounter().

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

gpu-compute/dispatcher.hh
gpu-compute/dispatcher.cc

Public Types

Public Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ Params

Constructor & Destructor Documentation

◆ GPUDispatcher()

◆ ~GPUDispatcher()

Member Function Documentation

◆ dispatch()

◆ exec()

◆ getOutstandingWbs()

◆ hsaTask()

◆ isReachingKernelEnd()

◆ notifyWgCompl()

◆ regStats()

◆ scheduleDispatch()

◆ serialize()

◆ setCommandProcessor()

◆ setShader()

◆ unserialize()

◆ updateInvCounter()

◆ updateWbCounter()

Member Data Documentation

◆ cyclesWaitingForDispatch

◆ dispatchActive

◆ doneIds

◆ execIds

◆ gpuCmdProc

◆ hsaQueueEntries

◆ numKernelLaunched

◆ shader

◆ tickEvent