#include <gpu_command_processor.hh>

Inheritance diagram for GPUCommandProcessor:

Classes
class	MQDDmaEvent
	Perform a DMA read of the MQD that corresponds to a hardware queue descriptor (HQD). More...

class	ReadDispIdOffsetDmaEvent
	Perform a DMA read of the read_dispatch_id_field_base_byte_offset field, which follows directly after the read_dispatch_id (the read pointer) in the amd_hsa_queue_t struct (aka memory queue descriptor (MQD)), to find the base address of the MQD. More...

Public Types
enum	AgentCmd { Nop = 0, Steal = 1 }

typedef GPUCommandProcessorParams	Params

Public Types inherited from HSADevice
typedef HSADeviceParams	Params

typedef std::function< void(const uint64_t &)>	HsaSignalCallbackFunction

Public Types inherited from DmaDevice
typedef DmaDeviceParams	Params

Public Types inherited from PioDevice
using	Params = PioDeviceParams

Public Types inherited from ClockedObject
using	Params = ClockedObjectParams
	Parameters of ClockedObject. More...

Public Types inherited from SimObject
typedef SimObjectParams	Params

Public Member Functions
	GPUCommandProcessor ()=delete

	GPUCommandProcessor (const Params &p)

void	setShader (Shader *shader)

Shader *	shader ()

void	submitAgentDispatchPkt (void *raw_pkt, uint32_t queue_id, Addr host_pkt_addr) override
	submitAgentDispatchPkt() is for accepting agent dispatch packets. More...

void	submitDispatchPkt (void *raw_pkt, uint32_t queue_id, Addr host_pkt_addr) override
	submitDispatchPkt() is the entry point into the CP from the HSAPP and is only meant to be used with AQL kernel dispatch packets. More...

void	submitVendorPkt (void *raw_pkt, uint32_t queue_id, Addr host_pkt_addr) override
	submitVendorPkt() is for accepting vendor-specific packets from the HSAPP. More...

void	attachDriver (HSADriver *driver) override

void	dispatchPkt (HSAQueueEntry *task)
	Once the CP has finished extracting all relevant information about a task and has initialized the ABI state, we send a description of the task to the dispatcher. More...

void	signalWakeupEvent (uint32_t event_id)

Tick	write (PacketPtr pkt) override
	Pure virtual function that the device must implement. More...

Tick	read (PacketPtr pkt) override
	Pure virtual function that the device must implement. More...

AddrRangeList	getAddrRanges () const override
	Every PIO device is obliged to provide an implementation that returns the address ranges the device responds to. More...

System *	system ()

void	updateHsaSignal (Addr signal_handle, uint64_t signal_value) override

uint64_t	functionalReadHsaSignal (Addr signal_handle) override

Addr	getHsaSignalValueAddr (Addr signal_handle)

Addr	getHsaSignalMailboxAddr (Addr signal_handle)

Addr	getHsaSignalEventAddr (Addr signal_handle)

Public Member Functions inherited from HSADevice
	HSADevice (const Params &p)

HSAPacketProcessor &	hsaPacketProc ()

void	dmaReadVirt (Addr host_addr, unsigned size, DmaCallback cb, void data, Tick delay=0)

void	dmaWriteVirt (Addr host_addr, unsigned size, DmaCallback cb, void data, Tick delay=0)

Public Member Functions inherited from DmaDevice
	DmaDevice (const Params &p)

virtual	~DmaDevice ()=default

void	dmaWrite (Addr addr, int size, Event event, uint8_t data, uint32_t sid, uint32_t ssid, Tick delay=0)

void	dmaWrite (Addr addr, int size, Event event, uint8_t data, Tick delay=0)

void	dmaRead (Addr addr, int size, Event event, uint8_t data, uint32_t sid, uint32_t ssid, Tick delay=0)

void	dmaRead (Addr addr, int size, Event event, uint8_t data, Tick delay=0)

bool	dmaPending () const

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

unsigned int	cacheBlockSize () const

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

Public Member Functions inherited from PioDevice
	PioDevice (const Params &p)

virtual	~PioDevice ()

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

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

Public Member Functions inherited from ClockedObject
	ClockedObject (const ClockedObjectParams &p)

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

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

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

	SimObject (const Params &p)

virtual	~SimObject ()

virtual const std::string	name () const

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

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...

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

	Group ()=delete

	Group (const Group &)=delete

Group &	operator= (const Group &)=delete

Public Member Functions inherited from Clocked
void	updateClockPeriod ()
	Update the tick to the current tick. More...

Tick	clockEdge (Cycles cycles=Cycles(0)) const
	Determine the tick when a cycle begins, by default the current one, but the argument also enables the caller to determine a future cycle. More...

Cycles	curCycle () const
	Determine the current cycle, corresponding to a tick aligned to a clock edge. More...

Tick	nextCycle () const
	Based on the clock of the object, determine the start tick of the first cycle that is at least one cycle in the future. More...

uint64_t	frequency () const

Tick	clockPeriod () const

double	voltage () const

Cycles	ticksToCycles (Tick t) const

Tick	cyclesToTicks (Cycles c) const

Private Member Functions
void	initABI (HSAQueueEntry *task)
	The CP is responsible for traversing all HSA-ABI-related data structures from memory and initializing the ABI state. More...

Private Attributes
Shader *	_shader

GPUDispatcher &	dispatcher

HSADriver *	driver

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)

Public Attributes inherited from ClockedObject
PowerState *	powerState

Protected Types inherited from HSADevice
typedef void(DmaDevice::*	DmaFnPtr) (Addr, int, Event , uint8_t , Tick)

Protected Member Functions inherited from HSADevice
void	dmaVirt (DmaFnPtr, Addr host_addr, unsigned size, DmaCallback cb, void data, Tick delay=0)

void	translateOrDie (Addr vaddr, Addr &paddr)
	HSADevices will perform DMA operations on VAs, and because page faults are not currently supported for HSADevices, we must be able to find the pages mapped for the process. More...

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 Member Functions inherited from Clocked
	Clocked (ClockDomain &clk_domain)
	Create a clocked object and set the clock domain based on the parameters. More...

	Clocked (Clocked &)=delete

Clocked &	operator= (Clocked &)=delete

virtual	~Clocked ()
	Virtual destructor due to inheritance. More...

void	resetClock () const
	Reset the object's clock using the current global tick value. More...

virtual void	clockPeriodUpdated ()
	A hook subclasses can implement so they can do any extra work that's needed when the clock rate is changed. More...

Protected Attributes inherited from HSADevice
HSAPacketProcessor *	hsaPP

Protected Attributes inherited from DmaDevice
DmaPort	dmaPort

Protected Attributes inherited from PioDevice
System *	sys

PioPort< PioDevice >	pioPort
	The pioPort that handles the requests for us and provides us requests that it sees. 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 57 of file gpu_command_processor.hh.

Member Typedef Documentation

◆ Params

typedef GPUCommandProcessorParams GPUCommandProcessor::Params

Definition at line 60 of file gpu_command_processor.hh.

Member Enumeration Documentation

◆ AgentCmd

enum GPUCommandProcessor::AgentCmd

Enumerator
Nop
Steal

Definition at line 68 of file gpu_command_processor.hh.

Constructor & Destructor Documentation

◆ GPUCommandProcessor() [1/2]

GPUCommandProcessor::GPUCommandProcessor ( )

delete

◆ GPUCommandProcessor() [2/2]

GPUCommandProcessor::GPUCommandProcessor ( const Params & p )

Definition at line 44 of file gpu_command_processor.cc.

References dispatcher, and GPUDispatcher::setCommandProcessor().

Member Function Documentation

◆ attachDriver()

void GPUCommandProcessor::attachDriver ( HSADriver * driver )

overridevirtual

Reimplemented from HSADevice.

Definition at line 195 of file gpu_command_processor.cc.

References driver, and fatal_if.

◆ dispatchPkt()

void GPUCommandProcessor::dispatchPkt ( HSAQueueEntry * task )

Once the CP has finished extracting all relevant information about a task and has initialized the ABI state, we send a description of the task to the dispatcher.

The dispatcher will create and dispatch WGs to the CUs.

Definition at line 280 of file gpu_command_processor.cc.

References GPUDispatcher::dispatch(), and dispatcher.

Referenced by GPUCommandProcessor::MQDDmaEvent::process().

◆ functionalReadHsaSignal()

uint64_t GPUCommandProcessor::functionalReadHsaSignal ( Addr signal_handle )

overridevirtual

Reimplemented from HSADevice.

Definition at line 151 of file gpu_command_processor.cc.

References getHsaSignalValueAddr(), system(), and System::threads.

Referenced by GPUDispatcher::notifyWgCompl().

◆ getAddrRanges()

AddrRangeList GPUCommandProcessor::getAddrRanges ( ) const

overridevirtual

Every PIO device is obliged to provide an implementation that returns the address ranges the device responds to.

Returns: a list of non-overlapping address ranges

Implements PioDevice.

Definition at line 317 of file gpu_command_processor.cc.

◆ getHsaSignalEventAddr()

Addr GPUCommandProcessor::getHsaSignalEventAddr ( Addr signal_handle )

inline

Definition at line 102 of file gpu_command_processor.hh.

Referenced by updateHsaSignal().

◆ getHsaSignalMailboxAddr()

Addr GPUCommandProcessor::getHsaSignalMailboxAddr ( Addr signal_handle )

inline

Definition at line 97 of file gpu_command_processor.hh.

Referenced by updateHsaSignal().

◆ getHsaSignalValueAddr()

Addr GPUCommandProcessor::getHsaSignalValueAddr ( Addr signal_handle )

inline

Definition at line 92 of file gpu_command_processor.hh.

Referenced by functionalReadHsaSignal(), and updateHsaSignal().

◆ initABI()

void GPUCommandProcessor::initABI ( HSAQueueEntry * task )

private

The CP is responsible for traversing all HSA-ABI-related data structures from memory and initializing the ABI state.

Information provided by the MQD, AQL packet, and code object metadata will be used to initialze register file state.

Definition at line 298 of file gpu_command_processor.cc.

References HSADevice::dmaReadVirt(), HSAPacketProcessor::getQueueDesc(), HSAQueueDescriptor::hostReadIndexPtr, HSADevice::hsaPP, and HSAQueueEntry::queueId().

Referenced by submitDispatchPkt().

◆ read()

Tick GPUCommandProcessor::read ( PacketPtr pkt )

inlineoverridevirtual

Pure virtual function that the device must implement.

Called when a read command is recieved by the port.

Parameters

pkt	Packet describing this request

Returns: number of ticks it took to complete

Implements PioDevice.

Definition at line 84 of file gpu_command_processor.hh.

◆ setShader()

void GPUCommandProcessor::setShader ( Shader * shader )

Definition at line 324 of file gpu_command_processor.cc.

References _shader, and shader().

◆ shader()

Shader * GPUCommandProcessor::shader ( )

Definition at line 330 of file gpu_command_processor.cc.

References _shader.

Referenced by setShader().

◆ signalWakeupEvent()

void GPUCommandProcessor::signalWakeupEvent ( uint32_t event_id )

Definition at line 286 of file gpu_command_processor.cc.

References driver, and HSADriver::signalWakeupEvent().

Referenced by updateHsaSignal().

◆ submitAgentDispatchPkt()

void GPUCommandProcessor::submitAgentDispatchPkt	(	void *	raw_pkt,
		uint32_t	queue_id,
		Addr	host_pkt_addr
	)

overridevirtual

submitAgentDispatchPkt() is for accepting agent dispatch packets.

These packets will control the dispatch of Wg on the device, and inform the host when a specified number of Wg have been executed on the device.

For now it simply finishes the pkt.

Reimplemented from HSADevice.

Definition at line 233 of file gpu_command_processor.cc.

References _hsa_agent_dispatch_packet_s::arg, HSAQueueEntry::completionSignal(), dispatcher, HSADevice::dmaWriteVirt(), DPRINTF, HSAPacketProcessor::finishPkt(), HSADevice::hsaPP, GPUDispatcher::hsaTask(), Nop(), panic, _hsa_agent_dispatch_packet_s::return_address, and _hsa_agent_dispatch_packet_s::type.

◆ submitDispatchPkt()

void GPUCommandProcessor::submitDispatchPkt	(	void *	raw_pkt,
		uint32_t	queue_id,
		Addr	host_pkt_addr
	)

overridevirtual

submitDispatchPkt() is the entry point into the CP from the HSAPP and is only meant to be used with AQL kernel dispatch packets.

After the HSAPP receives and extracts an AQL packet, it sends it to the CP, which is responsible for gathering all relevant information about a task, initializing CU state, and sending it to the dispatcher for WG creation and dispatch.

First we need capture all information from the the AQL pkt and the code object, then store it in an HSAQueueEntry. Once the packet and code are extracted, we extract information from the queue descriptor that the CP needs to perform state initialization on the CU. Finally we call dispatch() to send the task to the dispatcher. When the task completely finishes, we call finishPkt() on the HSA packet processor in order to remove the packet from the queue, and notify the runtime that the task has completed.

we need to read a pointer in the application's address space to pull out the kernel code descriptor.

The kernel_object is a pointer to the machine code, whose entry point is an 'amd_kernel_code_t' type, which is included in the kernel binary, and describes various aspects of the kernel. The desired entry is the 'kernel_code_entry_byte_offset' field, which provides the byte offset (positive or negative) from the address of the amd_kernel_code_t to the start of the machine instructions.

BLIT kernels don't have symbol names. BLIT kernels are built-in compute kernels issued by ROCm to handle DMAs for dGPUs when the SDMA hardware engines are unavailable or explicitly disabled. They can also be used to do copies that ROCm things would be better performed by the shader than the SDMA engines. They are also sometimes used on APUs to implement asynchronous memcopy operations from 2 pointers in host memory. I have no idea what BLIT stands for.

Reimplemented from HSADevice.

Definition at line 68 of file gpu_command_processor.cc.

◆ submitVendorPkt()

void GPUCommandProcessor::submitVendorPkt	(	void *	raw_pkt,
		uint32_t	queue_id,
		Addr	host_pkt_addr
	)

overridevirtual

submitVendorPkt() is for accepting vendor-specific packets from the HSAPP.

Vendor-specific packets may be used by the runtime to send commands to the HSA device that are specific to a particular vendor. The vendor-specific packets should be defined by the vendor in the runtime. TODO: For now we simply tell the HSAPP to finish the packet, however a future patch will update this method to provide the proper handling of any required vendor-specific packets. In the version of ROCm that is currently supported (1.6) the runtime will send packets that direct the CP to invalidate the GPUs caches. We do this automatically on each kernel launch in the CU, so this is safe for now.

Reimplemented from HSADevice.

Definition at line 219 of file gpu_command_processor.cc.

References HSAPacketProcessor::finishPkt(), and HSADevice::hsaPP.

◆ system()

System * GPUCommandProcessor::system ( )

Definition at line 311 of file gpu_command_processor.cc.

References PioDevice::sys.

Referenced by functionalReadHsaSignal(), and updateHsaSignal().

◆ updateHsaSignal()

void GPUCommandProcessor::updateHsaSignal	(	Addr	signal_handle,
		uint64_t	signal_value
	)

overridevirtual

Reimplemented from HSADevice.

Definition at line 160 of file gpu_command_processor.cc.

References HSADevice::dmaWriteVirt(), DPRINTF, getHsaSignalEventAddr(), getHsaSignalMailboxAddr(), getHsaSignalValueAddr(), signalWakeupEvent(), system(), and System::threads.

Referenced by GPUDispatcher::notifyWgCompl().

◆ write()

Tick GPUCommandProcessor::write ( PacketPtr pkt )

inlineoverridevirtual

Pure virtual function that the device must implement.

Called when a write command is recieved by the port.

Parameters

pkt	Packet describing this request

Returns: number of ticks it took to complete

Implements PioDevice.

Definition at line 83 of file gpu_command_processor.hh.

Member Data Documentation

◆ _shader

Shader* GPUCommandProcessor::_shader

private

Definition at line 108 of file gpu_command_processor.hh.

Referenced by setShader(), and shader().

◆ dispatcher

GPUDispatcher& GPUCommandProcessor::dispatcher

private

Definition at line 109 of file gpu_command_processor.hh.

Referenced by dispatchPkt(), GPUCommandProcessor(), and submitAgentDispatchPkt().

◆ driver

HSADriver* GPUCommandProcessor::driver

private

Definition at line 110 of file gpu_command_processor.hh.

Referenced by attachDriver(), and signalWakeupEvent().

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

gpu-compute/gpu_command_processor.hh
gpu-compute/gpu_command_processor.cc

Classes

Public Types

Public Member Functions

Private Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ Params

Member Enumeration Documentation

◆ AgentCmd

Constructor & Destructor Documentation

◆ GPUCommandProcessor() [1/2]

◆ GPUCommandProcessor() [2/2]

Member Function Documentation

◆ attachDriver()

◆ dispatchPkt()

◆ functionalReadHsaSignal()

◆ getAddrRanges()

◆ getHsaSignalEventAddr()

◆ getHsaSignalMailboxAddr()

◆ getHsaSignalValueAddr()

◆ initABI()

◆ read()

◆ setShader()

◆ shader()

◆ signalWakeupEvent()

◆ submitAgentDispatchPkt()

◆ submitDispatchPkt()

◆ submitVendorPkt()

◆ system()

◆ updateHsaSignal()

◆ write()

Member Data Documentation

◆ _shader

◆ dispatcher

◆ driver