#include <dispatcher.hh>

Inheritance diagram for gem5::GPUDispatcher:

Classes
struct	GPUDispatcherStats

Public Types
typedef GPUDispatcherParams	Params

Public Types inherited from gem5::SimObject
typedef SimObjectParams	Params

Public Member Functions
	GPUDispatcher (const Params &p)

	~GPUDispatcher ()

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

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

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)

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)

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

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.

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.

HSAQueueEntry *	hsaTask (int disp_id)

Public Member Functions inherited from gem5::SimObject
const Params &	params () const

	SimObject (const Params &p)

virtual	~SimObject ()

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

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

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

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

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

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

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

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

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

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

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

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

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

Public Member Functions inherited from gem5::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.

void	setCurTick (Tick newVal)

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

	EventManager (EventManager *em)

	EventManager (EventQueue *eq)

Public Member Functions inherited from gem5::Serializable
	Serializable ()

virtual	~Serializable ()

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

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

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

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

Public Member Functions inherited from gem5::Drainable
DrainState	drainState () const
	Return the current drain state of an object.

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

Public Member Functions inherited from gem5::statistics::Group
	Group (Group parent, const char name=nullptr)
	Construct a new statistics group.

virtual	~Group ()

virtual void	regStats ()
	Callback to set stat parameters.

virtual void	resetStats ()
	Callback to reset stats.

virtual void	preDumpStats ()
	Callback before stats are dumped.

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

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

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

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

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

void	mergeStatGroup (Group *block)
	Merge the contents (stats & children) of a block to this block.

	Group ()=delete

	Group (const Group &)=delete

Group &	operator= (const Group &)=delete

Public Member Functions inherited from gem5::Named
	Named (const std::string &name_)

virtual	~Named ()=default

virtual std::string	name () const

Protected Attributes
gem5::GPUDispatcher::GPUDispatcherStats	stats

Protected Attributes inherited from gem5::SimObject
const SimObjectParams &	_params
	Cached copy of the object parameters.

Protected Attributes inherited from gem5::EventManager
EventQueue *	eventq
	A pointer to this object's event queue.

Private Attributes
Shader *	shader

GPUCommandProcessor *	gpuCmdProc

EventFunctionWrapper	tickEvent

std::unordered_map< int, HSAQueueEntry * >	hsaQueueEntries

std::queue< int >	execIds

std::queue< int >	doneIds

bool	dispatchActive

bool	kernelExitEvents

Additional Inherited Members
Static Public Member Functions inherited from gem5::SimObject
static void	serializeAll (const std::string &cpt_dir)
	Create a checkpoint by serializing all SimObjects in the system.

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

static void	setSimObjectResolver (SimObjectResolver *resolver)
	There is a single object name resolver, and it is only set when simulation is restoring from checkpoints.

static SimObjectResolver *	getSimObjectResolver ()
	There is a single object name resolver, and it is only set when simulation is restoring from checkpoints.

Static Public Member Functions inherited from gem5::Serializable
static const std::string &	currentSection ()
	Gets the fully-qualified name of the active section.

static void	generateCheckpointOut (const std::string &cpt_dir, std::ofstream &outstream)
	Generate a checkpoint file so that the serialization can be routed to it.

Protected Member Functions inherited from gem5::Drainable
	Drainable ()

virtual	~Drainable ()

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

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

Detailed Description

Definition at line 62 of file dispatcher.hh.

Member Typedef Documentation

◆ Params

typedef GPUDispatcherParams gem5::GPUDispatcher::Params

Definition at line 65 of file dispatcher.hh.

Constructor & Destructor Documentation

◆ GPUDispatcher()

gem5::GPUDispatcher::GPUDispatcher ( const Params & p )

Definition at line 50 of file dispatcher.cc.

References exec().

◆ ~GPUDispatcher()

gem5::GPUDispatcher::~GPUDispatcher ( )

Definition at line 60 of file dispatcher.cc.

Member Function Documentation

◆ dispatch()

void gem5::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 116 of file dispatcher.cc.

References gem5::Clocked::clockPeriod(), gem5::curTick(), dispatchActive, gem5::HSAQueueEntry::dispatchId(), DPRINTF, execIds, hsaQueueEntries, gem5::HSAQueueEntry::kernelName(), gem5::GPUDispatcher::GPUDispatcherStats::numKernelLaunched, gem5::EventManager::schedule(), gem5::Event::scheduled(), shader, stats, and tickEvent.

Referenced by gem5::GPUCommandProcessor::dispatchPkt().

◆ exec()

void gem5::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 retry later

Definition at line 135 of file dispatcher.cc.

References gem5::curTick(), gem5::GPUDispatcher::GPUDispatcherStats::cyclesWaitingForDispatch, gem5::Shader::dispatchWorkgroups(), doneIds, DPRINTF, execIds, hsaQueueEntries, gem5::Shader::impl_kern_launch_acq, gem5::Shader::prepareInvalidate(), shader, stats, and gem5::Shader::updateContext().

Referenced by GPUDispatcher().

◆ getOutstandingWbs()

int gem5::GPUDispatcher::getOutstandingWbs ( int kern_id )

get kernel's outstanding cache writeback requests

Definition at line 282 of file dispatcher.cc.

References hsaQueueEntries.

Referenced by gem5::Shader::prepareFlush().

◆ hsaTask()

HSAQueueEntry * gem5::GPUDispatcher::hsaTask ( int disp_id )

Definition at line 65 of file dispatcher.cc.

References hsaQueueEntries.

Referenced by gem5::GPUCommandProcessor::submitAgentDispatchPkt().

◆ isReachingKernelEnd()

bool gem5::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 228 of file dispatcher.cc.

References hsaQueueEntries, and gem5::Wavefront::kernId.

Referenced by gem5::VegaISA::Inst_SOPP__S_ENDPGM::execute().

◆ notifyWgCompl()

void gem5::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.

Definition at line 297 of file dispatcher.cc.

References gem5::Clocked::clockPeriod(), gem5::Wavefront::computeUnit, gem5::ComputeUnit::cu_id, gem5::curTick(), DPRINTF, gem5::HSAPacketProcessor::finishPkt(), gpuCmdProc, gem5::GPUCommandProcessor::hsaPacketProc(), hsaQueueEntries, kernelExitEvents, gem5::Wavefront::kernId, gem5::Shader::requestKernelExitEvent(), gem5::EventManager::schedule(), gem5::Event::scheduled(), gem5::GPUCommandProcessor::sendCompletionSignal(), shader, tickEvent, and gem5::Wavefront::wgId.