#include <gpu_compute_driver.hh>

Inheritance diagram for gem5::GPUComputeDriver:

Classes
class	DriverWakeupEvent

class	EventList

class	EventTableEntry

Public Types
typedef GPUComputeDriverParams	Params

typedef class EventTableEntry	ETEntry

Public Types inherited from gem5::SimObject
typedef SimObjectParams	Params

Public Member Functions
	GPUComputeDriver (const Params &p)

int	ioctl (ThreadContext *tc, unsigned req, Addr ioc_buf) override
	Abstract method, invoked when the user program calls ioctl() on the file descriptor returned by a previous open(). More...

int	open (ThreadContext *tc, int mode, int flags) override
	Create an FD entry for the KFD inside of the owning process. More...

Addr	mmap (ThreadContext *tc, Addr start, uint64_t length, int prot, int tgt_flags, int tgt_fd, off_t offset) override
	Currently, mmap() will simply setup a mapping for the associated device's packet processor's doorbells and creates the event page. More...

virtual void	signalWakeupEvent (uint32_t event_id)

void	sleepCPU (ThreadContext *tc, uint32_t milliSecTimeout)

void	setMtype (RequestPtr req)
	Called by the compute units right before a request is issued to ruby. More...

int	doorbellSize ()

Public Member Functions inherited from gem5::EmulatedDriver
	EmulatedDriver (const EmulatedDriverParams &p)

bool	match (const std::string &s) const
	Check for a match with this driver's filename. More...

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. 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 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. 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 gem5::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 gem5::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 gem5::statistics::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 (statistics::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 gem5::Named
	Named (const std::string &name_)

virtual	~Named ()=default

virtual std::string	name () const

Private Types
enum	MtypeFlags { SHARED = 0 , READ_WRITE = 1 , CACHED = 2 , NUM_MTYPE_BITS }
	Mtype bits {Cached, Read Write, Shared} for caches. More...

Private Member Functions
void	registerUncacheableMemory (Addr start, Addr length)
	Register a region of host memory as uncacheable from the perspective of the dGPU. More...

Addr	gpuVmApeBase (int gpuNum) const
	The aperture (APE) base/limit pairs are set statically at startup by the real KFD. More...

Addr	gpuVmApeLimit (Addr apeBase) const

Addr	scratchApeBase (int gpuNum) const

Addr	scratchApeBaseV9 () const

Addr	scratchApeLimit (Addr apeBase) const

Addr	ldsApeBase (int gpuNum) const

Addr	ldsApeBaseV9 () const

Addr	ldsApeLimit (Addr apeBase) const

void	allocateGpuVma (Request::CacheCoherenceFlags mtype, Addr start, Addr length)
	Allocate/deallocate GPUVM VMAs for tracking virtual address allocations and properties on DGPUs. More...

Addr	deallocateGpuVma (Addr start)

void	allocateQueue (PortProxy &mem_proxy, Addr ioc_buf_addr)
	Forward relevant parameters to packet processor; queueId is used to link doorbell. More...

Private Attributes
GPUCommandProcessor *	device
	GPU that is controlled by this driver. More...

uint32_t	queueId

bool	isdGPU

GfxVersion	gfxVersion

int	dGPUPoolID

Addr	eventPage

uint32_t	eventSlotIndex

std::unordered_map< uint32_t, ETEntry >	ETable

AddrRangeMap< Request::CacheCoherenceFlags, 1 >	gpuVmas
	VMA structures for GPUVM memory. More...

Request::CacheCoherenceFlags	defaultMtype

std::unordered_map< ThreadContext *, EventList >	TCEvents

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. More...

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

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

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

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

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

Protected Member Functions inherited from gem5::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 gem5::EmulatedDriver
const std::string &	filename
	filename for opening this driver (under /dev) More...

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

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

Detailed Description

Definition at line 62 of file gpu_compute_driver.hh.

Member Typedef Documentation

◆ ETEntry

typedef class EventTableEntry gem5::GPUComputeDriver::ETEntry

Definition at line 86 of file gpu_compute_driver.hh.

◆ Params

typedef GPUComputeDriverParams gem5::GPUComputeDriver::Params

Definition at line 65 of file gpu_compute_driver.hh.

Member Enumeration Documentation

◆ MtypeFlags

enum gem5::GPUComputeDriver::MtypeFlags

private

Mtype bits {Cached, Read Write, Shared} for caches.

Enumerator
SHARED
READ_WRITE
CACHED
NUM_MTYPE_BITS

Definition at line 167 of file gpu_compute_driver.hh.

Constructor & Destructor Documentation

◆ GPUComputeDriver()

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

Definition at line 60 of file gpu_compute_driver.cc.

References gem5::GPUCommandProcessor::attachDriver(), gem5::Request::CACHED, defaultMtype, device, DPRINTF, gem5::VegaISA::p, gem5::Request::READ_WRITE, gem5::Flags< T >::set(), and gem5::Request::SHARED.

Member Function Documentation

◆ allocateGpuVma()

void gem5::GPUComputeDriver::allocateGpuVma	(	Request::CacheCoherenceFlags	mtype,
		Addr	start,
		Addr	length
	)

private

Allocate/deallocate GPUVM VMAs for tracking virtual address allocations and properties on DGPUs.

For now, we use these to track MTYPE and to be able to select which pages to unmap when the user provides us with a handle during the free ioctl.

Definition at line 997 of file gpu_compute_driver.cc.

References DPRINTF, gem5::AddrRange::end(), fatal_if, gpuVmas, and gem5::AddrRange::start().

Referenced by ioctl().

◆ allocateQueue()

void gem5::GPUComputeDriver::allocateQueue	(	PortProxy &	mem_proxy,
		Addr	ioc_buf
	)

private

Forward relevant parameters to packet processor; queueId is used to link doorbell.

The queueIDs are not re-used in current implementation, and we allocate only one page (4096 bytes) for doorbells, so check if this queueID can be mapped into that page.

Definition at line 159 of file gpu_compute_driver.cc.

References gem5::BaseBufferArg::copyIn(), gem5::BaseBufferArg::copyOut(), device, doorbellSize(), fatal, gfxVersion, gem5::GPUCommandProcessor::hsaPacketProc(), KFD_MMAP_GPU_ID, KFD_MMAP_TYPE_DOORBELL, gem5::Named::name(), PAGE_SHIFT, queueId, and gem5::HSAPacketProcessor::setDeviceQueueDesc().

Referenced by ioctl().

◆ deallocateGpuVma()

Addr gem5::GPUComputeDriver::deallocateGpuVma ( Addr start )

private

Definition at line 1009 of file gpu_compute_driver.cc.

References DPRINTF, and gpuVmas.

Referenced by ioctl().

◆ doorbellSize()

int gem5::GPUComputeDriver::doorbellSize ( )

inline

Definition at line 86 of file gpu_compute_driver.hh.

References fatal, and gfxVersion.

Referenced by allocateQueue(), ioctl(), and gem5::HSAPacketProcessor::write().

◆ gpuVmApeBase()

Addr gem5::GPUComputeDriver::gpuVmApeBase ( int gpuNum ) const

private

The aperture (APE) base/limit pairs are set statically at startup by the real KFD.

AMD x86_64 CPUs only use the areas in the 64b address space where VA[63:47] == 0x1ffff or VA[63:47] = 0. These methods generate the APE base/limit pairs in exactly the same way as the real KFD does, which ensures these APEs do not fall into the CPU's address space

see the macros in the KFD driver in the ROCm Linux kernel source:

drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c

Definition at line 945 of file gpu_compute_driver.cc.

Referenced by ioctl().

◆ gpuVmApeLimit()

Addr gem5::GPUComputeDriver::gpuVmApeLimit ( Addr apeBase ) const

private

Definition at line 951 of file gpu_compute_driver.cc.

Referenced by ioctl().

◆ ioctl()

int gem5::GPUComputeDriver::ioctl	(	ThreadContext *	tc,
		unsigned	req,
		Addr	buf
	)

overridevirtual

Abstract method, invoked when the user program calls ioctl() on the file descriptor returned by a previous open().

The parameters are the same as those passed in to ioctlFunc() (q.v.).

Returns: The return code for the ioctl, or the negation of the errno (see the SyscallReturn class).

This is where the runtime requests MTYPE from an aperture. Basically, the globally memory aperture is divided up into a default aperture and an alternate aperture each of which have their own MTYPE policies. This is done to mark a small piece of the global memory as uncacheable. Host memory mappings will be carved out of this uncacheable aperture, which is how they implement 'coherent' host/device memory on dGPUs.

TODO: Need to reflect per-aperture MTYPE policies based on this call.

Derive all clock counters based on the tick. All device clocks are identical and perfectly in sync.

Set the GPUVM/LDS/Scratch APEs exactly as they are in the real driver, see the KFD driver in the ROCm Linux kernel source: drivers/gpu/drm/amd/amdkfd/kfd_flat_memory.c

While the GPU node numbers start at 0, we add 1 to force the count to start at 1. This is to ensure that the base/limit addresses are calculated correctly.

The CPU's 64b address space can only use the areas with VA[63:47] == 0x1ffff or VA[63:47] == 0, therefore we must ensure that the apertures do not fall in the CPU's address space.

In real hardware, this IOCTL maps host memory, dGPU memory, or dGPU doorbells into GPUVM space. Essentially, ROCm implements SVM by carving out a region of free VA space that both the host and GPUVM can agree upon. The entire GPU VA space is reserved on the host using a fixed mmap at a low VA range that is also directly accessable by the GPU's limited number of VA bits. When we actually call memory allocation later in the program, this IOCTL is invoked to create BOs/VMAs in the driver and bind them to physical memory/doorbells.

For gem5, we don't need to carve out any GPUVM space here (we don't support GPUVM and use host page tables on the GPU directly). We can can just use the existing host SVM region. We comment on each memory type seperately.

Called to map an already allocated region of memory to this GPU's GPUVM VA space. We don't need to implement this in the simulator since we only have a single VM system. If the region has already been allocated somewhere like the CPU, then it's already visible to the device.

Implements gem5::EmulatedDriver.

Definition at line 229 of file gpu_compute_driver.cc.

References allocateGpuVma(), allocateQueue(), AMDKFD_IOC_ACQUIRE_VM, AMDKFD_IOC_ALLOC_MEMORY_OF_GPU, AMDKFD_IOC_ALLOC_QUEUE_GWS, AMDKFD_IOC_CREATE_EVENT, AMDKFD_IOC_CREATE_QUEUE, AMDKFD_IOC_DBG_ADDRESS_WATCH, AMDKFD_IOC_DBG_REGISTER, AMDKFD_IOC_DBG_UNREGISTER, AMDKFD_IOC_DBG_WAVE_CONTROL, AMDKFD_IOC_DESTROY_EVENT, AMDKFD_IOC_DESTROY_QUEUE, AMDKFD_IOC_FREE_MEMORY_OF_GPU, AMDKFD_IOC_GET_CLOCK_COUNTERS, AMDKFD_IOC_GET_DMABUF_INFO, AMDKFD_IOC_GET_PROCESS_APERTURES, AMDKFD_IOC_GET_PROCESS_APERTURES_NEW, AMDKFD_IOC_GET_QUEUE_WAVE_STATE, AMDKFD_IOC_GET_TILE_CONFIG, AMDKFD_IOC_GET_VERSION, AMDKFD_IOC_IMPORT_DMABUF, AMDKFD_IOC_MAP_MEMORY_TO_GPU, AMDKFD_IOC_RESET_EVENT, AMDKFD_IOC_SET_CU_MASK, AMDKFD_IOC_SET_EVENT, AMDKFD_IOC_SET_MEMORY_POLICY, AMDKFD_IOC_SET_SCRATCH_BACKING_VA, AMDKFD_IOC_SET_TRAP_HANDLER, AMDKFD_IOC_SMI_EVENTS, AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU, AMDKFD_IOC_UPDATE_QUEUE, AMDKFD_IOC_WAIT_EVENTS, gem5::Flags< T >::clear(), gem5::BaseBufferArg::copyIn(), gem5::BaseBufferArg::copyOut(), gem5::ThreadContext::cpuId(), gem5::curTick(), deallocateGpuVma(), defaultMtype, device, dGPUPoolID, gem5::divCeil(), doorbellSize(), DPRINTF, ETable, eventSlotIndex, fatal, fatal_if, gem5::BaseMMU::flushAll(), gem5::FullSystem, gem5::ThreadContext::getMMUPtr(), gem5::ThreadContext::getProcessPtr(), gfxVersion, gpuVmApeBase(), gpuVmApeLimit(), gem5::GPUCommandProcessor::hsaPacketProc(), gem5::ArmISA::i, isdGPU, KFD_IOC_ALLOC_MEM_FLAGS_COHERENT, KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL, KFD_IOC_ALLOC_MEM_FLAGS_GTT, KFD_IOC_ALLOC_MEM_FLAGS_USERPTR, KFD_IOC_ALLOC_MEM_FLAGS_VRAM, KFD_IOC_EVENT_SIGNAL, KFD_IOCTL_MAJOR_VERSION, KFD_IOCTL_MINOR_VERSION, KFD_MMAP_TYPE_EVENTS, ldsApeBase(), ldsApeBaseV9(), ldsApeLimit(), gem5::Process::memState, gem5::sim_clock::as_int::ns, PAGE_SHIFT, gem5::X86ISA::PageBytes, panic_if, gem5::HSAPacketProcessor::pioAddr, queueId, scratchApeBase(), scratchApeBaseV9(), scratchApeLimit(), signalWakeupEvent(), gem5::BaseBufferArg::size, sleepCPU(), SLOTS_PER_PAGE, TCEvents, gem5::HSAPacketProcessor::unsetDeviceQueueDesc(), and warn.

◆ ldsApeBase()

Addr gem5::GPUComputeDriver::ldsApeBase ( int gpuNum ) const

private

Definition at line 977 of file gpu_compute_driver.cc.

Referenced by ioctl().

◆ ldsApeBaseV9()

Addr gem5::GPUComputeDriver::ldsApeBaseV9 ( ) const

private

Definition at line 985 of file gpu_compute_driver.cc.

Referenced by ioctl().

◆ ldsApeLimit()

Addr gem5::GPUComputeDriver::ldsApeLimit ( Addr apeBase ) const

private

Definition at line 991 of file gpu_compute_driver.cc.

Referenced by ioctl().

◆ mmap()

Addr gem5::GPUComputeDriver::mmap	(	ThreadContext *	tc,
		Addr	start,
		uint64_t	length,
		int	prot,
		int	tgt_flags,
		int	tgt_fd,
		off_t	offset
	)

overridevirtual

Currently, mmap() will simply setup a mapping for the associated device's packet processor's doorbells and creates the event page.

We don't actually access these pages. We just need to reserve some VA space. See commit id 5ce8abce for details on how events are currently implemented.

Reimplemented from gem5::EmulatedDriver.

Definition at line 108 of file gpu_compute_driver.cc.

References device, DPRINTF, eventPage, gem5::ThreadContext::getProcessPtr(), gem5::GPUCommandProcessor::hsaPacketProc(), KFD_MMAP_TYPE_DOORBELL, KFD_MMAP_TYPE_EVENTS, KFD_MMAP_TYPE_MASK, KFD_SIGNAL_EVENT_LIMIT, gem5::Process::memState, gem5::ArmISA::offset, PAGE_SHIFT, panic_if, gem5::HSAPacketProcessor::pioAddr, and warn_once.

◆ open()

int gem5::GPUComputeDriver::open	(	ThreadContext *	tc,
		int	mode,
		int	flags
	)

overridevirtual

Create an FD entry for the KFD inside of the owning process.

Implements gem5::EmulatedDriver.

Definition at line 94 of file gpu_compute_driver.cc.

References DPRINTF, gem5::EmulatedDriver::filename, and gem5::ThreadContext::getProcessPtr().

◆ registerUncacheableMemory()

void gem5::GPUComputeDriver::registerUncacheableMemory	(	Addr	start,
		Addr	length
	)

private

Register a region of host memory as uncacheable from the perspective of the dGPU.

◆ scratchApeBase()

Addr gem5::GPUComputeDriver::scratchApeBase ( int gpuNum ) const

private

Definition at line 957 of file gpu_compute_driver.cc.

Referenced by ioctl().

◆ scratchApeBaseV9()

Addr gem5::GPUComputeDriver::scratchApeBaseV9 ( ) const

private

Definition at line 965 of file gpu_compute_driver.cc.

Referenced by ioctl().

◆ scratchApeLimit()

Addr gem5::GPUComputeDriver::scratchApeLimit ( Addr apeBase ) const

private

Definition at line 971 of file gpu_compute_driver.cc.

Referenced by ioctl().

◆ setMtype()

void gem5::GPUComputeDriver::setMtype ( RequestPtr req )

Called by the compute units right before a request is issued to ruby.

This uses our VMAs to correctly set the MTYPE on a per-request basis. In real hardware, this is actually done through PTE bits in GPUVM. Since we are running a single VM (x86 PT) system, the MTYPE bits aren't available. Adding GPUVM specific bits to x86 page tables probably isn't the best way to proceed. For now we just have the driver set these until we implement a proper dual PT system.

Definition at line 1022 of file gpu_compute_driver.cc.

References defaultMtype, DPRINTF, gem5::AddrRange::end(), gem5::FullSystem, gpuVmas, isdGPU, gem5::RangeSize(), and gem5::AddrRange::start().

Referenced by gem5::ComputeUnit::sendRequest().

◆ signalWakeupEvent()

void gem5::GPUComputeDriver::signalWakeupEvent ( uint32_t event_id )

virtual

Definition at line 191 of file gpu_compute_driver.cc.

References DPRINTF, ETable, eventSlotIndex, panic_if, and TCEvents.

Referenced by ioctl(), and gem5::GPUCommandProcessor::signalWakeupEvent().

◆ sleepCPU()

void gem5::GPUComputeDriver::sleepCPU	(	ThreadContext *	tc,
		uint32_t	milliSecTimeout
	)

Definition at line 933 of file gpu_compute_driver.cc.

References gem5::ThreadContext::cpuId(), DPRINTF, gem5::ThreadContext::suspend(), and TCEvents.

Referenced by ioctl().

Member Data Documentation

◆ defaultMtype

Request::CacheCoherenceFlags gem5::GPUComputeDriver::defaultMtype

private

Definition at line 175 of file gpu_compute_driver.hh.

Referenced by GPUComputeDriver(), ioctl(), and setMtype().

◆ device

GPUCommandProcessor* gem5::GPUComputeDriver::device

private

GPU that is controlled by this driver.

Definition at line 149 of file gpu_compute_driver.hh.

Referenced by allocateQueue(), GPUComputeDriver(), ioctl(), and mmap().

◆ dGPUPoolID

int gem5::GPUComputeDriver::dGPUPoolID

private

Definition at line 153 of file gpu_compute_driver.hh.

Referenced by ioctl().

◆ ETable

std::unordered_map<uint32_t, ETEntry> gem5::GPUComputeDriver::ETable

private

Definition at line 157 of file gpu_compute_driver.hh.

Referenced by gem5::GPUComputeDriver::EventList::clearEvents(), ioctl(), and signalWakeupEvent().

◆ eventPage

Addr gem5::GPUComputeDriver::eventPage

private

Definition at line 154 of file gpu_compute_driver.hh.

Referenced by mmap().

◆ eventSlotIndex

uint32_t gem5::GPUComputeDriver::eventSlotIndex

private

Definition at line 155 of file gpu_compute_driver.hh.

Referenced by gem5::GPUComputeDriver::EventList::clearEvents(), ioctl(), and signalWakeupEvent().

◆ gfxVersion

GfxVersion gem5::GPUComputeDriver::gfxVersion

private

Definition at line 152 of file gpu_compute_driver.hh.

Referenced by allocateQueue(), doorbellSize(), and ioctl().

◆ gpuVmas

AddrRangeMap<Request::CacheCoherenceFlags, 1> gem5::GPUComputeDriver::gpuVmas

private

VMA structures for GPUVM memory.

Definition at line 162 of file gpu_compute_driver.hh.

Referenced by allocateGpuVma(), deallocateGpuVma(), and setMtype().

◆ isdGPU

bool gem5::GPUComputeDriver::isdGPU

private

Definition at line 151 of file gpu_compute_driver.hh.

Referenced by ioctl(), and setMtype().

◆ queueId

uint32_t gem5::GPUComputeDriver::queueId

private

Definition at line 150 of file gpu_compute_driver.hh.

Referenced by allocateQueue(), and ioctl().

◆ TCEvents

std::unordered_map<ThreadContext *, EventList> gem5::GPUComputeDriver::TCEvents

private

Definition at line 206 of file gpu_compute_driver.hh.

Referenced by ioctl(), signalWakeupEvent(), and sleepCPU().

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

gpu-compute/gpu_compute_driver.hh
gpu-compute/gpu_compute_driver.cc

Classes

Public Types

Public Member Functions

Private Types

Private Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ ETEntry

◆ Params

Member Enumeration Documentation

◆ MtypeFlags

Constructor & Destructor Documentation

◆ GPUComputeDriver()

Member Function Documentation

◆ allocateGpuVma()

◆ allocateQueue()

◆ deallocateGpuVma()

◆ doorbellSize()

◆ gpuVmApeBase()

◆ gpuVmApeLimit()

◆ ioctl()

◆ ldsApeBase()

◆ ldsApeBaseV9()

◆ ldsApeLimit()

◆ mmap()

◆ open()

◆ registerUncacheableMemory()

◆ scratchApeBase()

◆ scratchApeBaseV9()

◆ scratchApeLimit()

◆ setMtype()

◆ signalWakeupEvent()

◆ sleepCPU()

Member Data Documentation

◆ defaultMtype

◆ device

◆ dGPUPoolID

◆ ETable

◆ eventPage

◆ eventSlotIndex

◆ gfxVersion

◆ gpuVmas

◆ isdGPU

◆ queueId

◆ TCEvents