#include "Target/AMDGPU/AMDGPUSubtarget.h"

Inheritance diagram for llvm::AMDGPUSubtarget:

Public Types
enum	Generation { INVALID = 0 , R600 = 1 , R700 = 2 , EVERGREEN = 3 , NORTHERN_ISLANDS = 4 , SOUTHERN_ISLANDS = 5 , SEA_ISLANDS = 6 , VOLCANIC_ISLANDS = 7 , GFX9 = 8 , GFX10 = 9 , GFX11 = 10 , GFX12 = 11 }

Public Member Functions
	AMDGPUSubtarget (Triple TT)

std::pair< unsigned, unsigned >	getDefaultFlatWorkGroupSize (CallingConv::ID CC) const

std::pair< unsigned, unsigned >	getFlatWorkGroupSizes (const Function &F) const

std::optional< unsigned >	getReqdWorkGroupSize (const Function &F, unsigned Dim) const

bool	hasWavefrontsEvenlySplittingXDim (const Function &F, bool REquiresUniformYZ=false) const

std::pair< unsigned, unsigned >	getWavesPerEU (const Function &F) const

std::pair< unsigned, unsigned >	getWavesPerEU (const Function &F, std::pair< unsigned, unsigned > FlatWorkGroupSizes) const
	Overload which uses the specified values for the flat work group sizes, rather than querying the function itself.

std::pair< unsigned, unsigned >	getWavesPerEU (std::pair< unsigned, unsigned > FlatWorkGroupSizes, unsigned LDSBytes, const Function &F) const
	Overload which uses the specified values for the flat workgroup sizes and LDS space rather than querying the function itself.

std::pair< unsigned, unsigned >	getEffectiveWavesPerEU (std::pair< unsigned, unsigned > RequestedWavesPerEU, std::pair< unsigned, unsigned > FlatWorkGroupSizes, unsigned LDSBytes) const
	Returns the target minimum/maximum number of waves per EU.

unsigned	getMaxLocalMemSizeWithWaveCount (unsigned WaveCount, const Function &) const
	Return the amount of LDS that can be used that will not restrict the occupancy lower than WaveCount.

std::pair< unsigned, unsigned >	getOccupancyWithWorkGroupSizes (uint32_t LDSBytes, const Function &F) const
	Subtarget's minimum/maximum occupancy, in number of waves per EU, that can be achieved when the only function running on a CU is `F` and each workgroup running the function requires `LDSBytes` bytes of LDS space.

std::pair< unsigned, unsigned >	getOccupancyWithWorkGroupSizes (uint32_t LDSBytes, std::pair< unsigned, unsigned > FlatWorkGroupSizes) const
	Overload which uses the specified values for the flat work group sizes, rather than querying the function itself.

std::pair< unsigned, unsigned >	getOccupancyWithWorkGroupSizes (const MachineFunction &MF) const
	Subtarget's minimum/maximum occupancy, in number of waves per EU, that can be achieved when the only function running on a CU is `MF`.

bool	isAmdHsaOS () const

bool	isAmdPalOS () const

bool	isMesa3DOS () const

bool	isMesaKernel (const Function &F) const

bool	isAmdHsaOrMesa (const Function &F) const

bool	isGCN () const

bool	isGCN3Encoding () const

bool	has16BitInsts () const

bool	hasTrue16BitInsts () const
	Return true if the subtarget supports True16 instructions.

bool	useRealTrue16Insts () const
	Return true if real (non-fake) variants of True16 instructions using 16-bit registers should be code-generated.

bool	hasBF16TransInsts () const

bool	hasBF16ConversionInsts () const

bool	hasBF16PackedInsts () const

bool	hasMadMixInsts () const

bool	hasFP8ConversionScaleInsts () const

bool	hasBF8ConversionScaleInsts () const

bool	hasFP4ConversionScaleInsts () const

bool	hasFP6BF6ConversionScaleInsts () const

bool	hasF16BF16ToFP6BF6ConversionScaleInsts () const

bool	hasCvtPkF16F32Inst () const

bool	hasF32ToF16BF16ConversionSRInsts () const

bool	hasMadMacF32Insts () const

bool	hasDsSrc2Insts () const

bool	hasSDWA () const

bool	hasVOP3PInsts () const

bool	hasMulI24 () const

bool	hasMulU24 () const

bool	hasSMulHi () const

bool	hasInv2PiInlineImm () const

bool	hasFminFmaxLegacy () const

bool	hasTrigReducedRange () const

bool	hasFastFMAF32 () const

bool	isPromoteAllocaEnabled () const

unsigned	getWavefrontSize () const

unsigned	getWavefrontSizeLog2 () const

unsigned	getLocalMemorySize () const
	Return the maximum number of bytes of LDS available for all workgroups running on the same WGP or CU.

unsigned	getAddressableLocalMemorySize () const
	Return the maximum number of bytes of LDS that can be allocated to a single workgroup.

unsigned	getEUsPerCU () const
	Number of SIMDs/EUs (execution units) per "CU" ("compute unit"), where the "CU" is the unit onto which workgroups are mapped.

Align	getAlignmentForImplicitArgPtr () const

unsigned	getExplicitKernelArgOffset () const
	Returns the offset in bytes from the start of the input buffer of the first explicit kernel argument.

virtual unsigned	getMaxWorkGroupsPerCU (unsigned FlatWorkGroupSize) const =0

virtual unsigned	getMinFlatWorkGroupSize () const =0

virtual unsigned	getMaxFlatWorkGroupSize () const =0

virtual unsigned	getWavesPerEUForWorkGroup (unsigned FlatWorkGroupSize) const =0

virtual unsigned	getMinWavesPerEU () const =0

unsigned	getMaxWavesPerEU () const

unsigned	getMaxWorkitemID (const Function &Kernel, unsigned Dimension) const
	Return the maximum workitem ID value in the function, for the given (0, 1, 2) dimension.

SmallVector< unsigned >	getMaxNumWorkGroups (const Function &F) const
	Return the number of work groups for the function.

bool	isSingleLaneExecution (const Function &Kernel) const
	Return true if only a single workitem can be active in a wave.

bool	makeLIDRangeMetadata (Instruction *I) const
	Creates value range metadata on an workitemid.* intrinsic call or load.

unsigned	getImplicitArgNumBytes (const Function &F) const

uint64_t	getExplicitKernArgSize (const Function &F, Align &MaxAlign) const

unsigned	getKernArgSegmentSize (const Function &F, Align &MaxAlign) const

AMDGPUDwarfFlavour	getAMDGPUDwarfFlavour () const

virtual	~AMDGPUSubtarget ()=default

Static Public Member Functions
static const AMDGPUSubtarget &	get (const MachineFunction &MF)

static const AMDGPUSubtarget &	get (const TargetMachine &TM, const Function &F)

Protected Attributes
bool	GCN3Encoding = false

bool	Has16BitInsts = false

bool	HasTrue16BitInsts = false

bool	HasFP8ConversionScaleInsts = false

bool	HasBF8ConversionScaleInsts = false

bool	HasFP4ConversionScaleInsts = false

bool	HasFP6BF6ConversionScaleInsts = false

bool	HasF16BF16ToFP6BF6ConversionScaleInsts = false

bool	HasCvtPkF16F32Inst = false

bool	HasF32ToF16BF16ConversionSRInsts = false

bool	EnableRealTrue16Insts = false

bool	HasBF16TransInsts = false

bool	HasBF16ConversionInsts = false

bool	HasBF16PackedInsts = false

bool	HasMadMixInsts = false

bool	HasMadMacF32Insts = false

bool	HasDsSrc2Insts = false

bool	HasSDWA = false

bool	HasVOP3PInsts = false

bool	HasMulI24 = true

bool	HasMulU24 = true

bool	HasSMulHi = false

bool	HasInv2PiInlineImm = false

bool	HasFminFmaxLegacy = true

bool	EnablePromoteAlloca = false

bool	HasTrigReducedRange = false

bool	FastFMAF32 = false

unsigned	EUsPerCU = 4

unsigned	MaxWavesPerEU = 10

unsigned	LocalMemorySize = 0

unsigned	AddressableLocalMemorySize = 0

char	WavefrontSizeLog2 = 0

Detailed Description

Definition at line 30 of file AMDGPUSubtarget.h.

Member Enumeration Documentation

◆ Generation

enum llvm::AMDGPUSubtarget::Generation

Enumerator
INVALID
R600
R700
EVERGREEN
NORTHERN_ISLANDS
SOUTHERN_ISLANDS
SEA_ISLANDS
VOLCANIC_ISLANDS
GFX9
GFX10
GFX11
GFX12

Definition at line 32 of file AMDGPUSubtarget.h.

Constructor & Destructor Documentation

◆ AMDGPUSubtarget()

AMDGPUSubtarget::AMDGPUSubtarget ( Triple TT )

Definition at line 35 of file AMDGPUSubtarget.cpp.

◆ ~AMDGPUSubtarget()

virtual llvm::AMDGPUSubtarget::~AMDGPUSubtarget ( )

virtualdefault

Member Function Documentation

◆ get() [1/2]

const AMDGPUSubtarget & AMDGPUSubtarget::get ( const MachineFunction & MF )

static

Definition at line 419 of file AMDGPUSubtarget.cpp.

References llvm::MachineFunction::getSubtarget(), llvm::MachineFunction::getTarget(), llvm::TargetMachine::getTargetTriple(), and llvm::Triple::isAMDGCN().

Referenced by llvm::AMDGPUTargetLowering::analyzeFormalArgumentsCompute().

◆ get() [2/2]

const AMDGPUSubtarget & AMDGPUSubtarget::get	(	const TargetMachine &	TM,
		const Function &	F
	)

static

Definition at line 425 of file AMDGPUSubtarget.cpp.

References F.

◆ getAddressableLocalMemorySize()

unsigned llvm::AMDGPUSubtarget::getAddressableLocalMemorySize ( ) const

inline

Return the maximum number of bytes of LDS that can be allocated to a single workgroup.

For GFX10-GFX12 in WGP mode this is limited to 64k even though the WGP has 128k in total.

Definition at line 327 of file AMDGPUSubtarget.h.

References AddressableLocalMemorySize.

◆ getAlignmentForImplicitArgPtr()

Align llvm::AMDGPUSubtarget::getAlignmentForImplicitArgPtr ( ) const

inline

Definition at line 336 of file AMDGPUSubtarget.h.

References isAmdHsaOS().

Referenced by llvm::SITargetLowering::allocatePreloadKernArgSGPRs(), llvm::AMDGPUTargetLowering::getImplicitParameterOffset(), and getKernArgSegmentSize().

◆ getAMDGPUDwarfFlavour()

AMDGPUDwarfFlavour AMDGPUSubtarget::getAMDGPUDwarfFlavour ( ) const

Returns: Corresponding DWARF register number mapping flavour for the WavefrontSize.

Definition at line 414 of file AMDGPUSubtarget.cpp.

References getWavefrontSize(), llvm::Wave32, and llvm::Wave64.

◆ getDefaultFlatWorkGroupSize()

std::pair< unsigned, unsigned > AMDGPUSubtarget::getDefaultFlatWorkGroupSize ( CallingConv::ID CC ) const

Returns: Default range flat work group size for a calling convention.

Definition at line 145 of file AMDGPUSubtarget.cpp.

References llvm::CallingConv::AMDGPU_ES, llvm::CallingConv::AMDGPU_GS, llvm::CallingConv::AMDGPU_HS, llvm::CallingConv::AMDGPU_LS, llvm::CallingConv::AMDGPU_PS, llvm::CallingConv::AMDGPU_VS, getMaxFlatWorkGroupSize(), and getWavefrontSize().

Referenced by getFlatWorkGroupSizes().

◆ getEffectiveWavesPerEU()

std::pair< unsigned, unsigned > AMDGPUSubtarget::getEffectiveWavesPerEU	(	std::pair< unsigned, unsigned >	RequestedWavesPerEU,
		std::pair< unsigned, unsigned >	FlatWorkGroupSizes,
		unsigned	LDSBytes
	)		const

Returns the target minimum/maximum number of waves per EU.

This is based on the minimum/maximum number of RequestedWavesPerEU and further limited by the maximum achievable occupancy derived from the range of FlatWorkGroupSizes and number of LDSBytes per workgroup.

Definition at line 182 of file AMDGPUSubtarget.cpp.

References llvm::Default, getMaxWavesPerEU(), getOccupancyWithWorkGroupSizes(), and getWavesPerEUForWorkGroup().

Referenced by getWavesPerEU().

◆ getEUsPerCU()

unsigned llvm::AMDGPUSubtarget::getEUsPerCU ( ) const

inline

Number of SIMDs/EUs (execution units) per "CU" ("compute unit"), where the "CU" is the unit onto which workgroups are mapped.

This takes WGP mode vs. CU mode into account.

Definition at line 334 of file AMDGPUSubtarget.h.

References EUsPerCU.

Referenced by getMaxLocalMemSizeWithWaveCount(), and getOccupancyWithWorkGroupSizes().

◆ getExplicitKernArgSize()

uint64_t AMDGPUSubtarget::getExplicitKernArgSize	(	const Function &	F,
		Align &	MaxAlign
	)		const

Definition at line 367 of file AMDGPUSubtarget.cpp.

References llvm::alignTo(), llvm::CallingConv::AMDGPU_KERNEL, assert(), DL, F, and llvm::CallingConv::SPIR_KERNEL.

Referenced by getKernArgSegmentSize().

◆ getExplicitKernelArgOffset()

unsigned llvm::AMDGPUSubtarget::getExplicitKernelArgOffset ( ) const

inline

Returns the offset in bytes from the start of the input buffer of the first explicit kernel argument.

Definition at line 342 of file AMDGPUSubtarget.h.

References llvm::Triple::AMDHSA, llvm::Triple::AMDPAL, llvm::Triple::getOS(), llvm_unreachable, llvm::Triple::Mesa3D, and llvm::Triple::UnknownOS.

Referenced by llvm::SITargetLowering::allocatePreloadKernArgSGPRs(), llvm::AMDGPUTargetLowering::getImplicitParameterOffset(), getKernArgSegmentSize(), and llvm::AMDGPUCallLowering::lowerFormalArgumentsKernel().

◆ getFlatWorkGroupSizes()

std::pair< unsigned, unsigned > AMDGPUSubtarget::getFlatWorkGroupSizes ( const Function & F ) const

Returns: Subtarget's default pair of minimum/maximum flat work group sizes for function F, or minimum/maximum flat work group sizes explicitly requested using "amdgpu-flat-work-group-size" attribute attached to function F.; Subtarget's default values if explicitly requested values cannot be converted to integer, or violate subtarget's specifications.

Definition at line 159 of file AMDGPUSubtarget.cpp.

References llvm::Default, F, getDefaultFlatWorkGroupSize(), llvm::AMDGPU::getIntegerPairAttribute(), getMaxFlatWorkGroupSize(), and getMinFlatWorkGroupSize().

Referenced by llvm::GCNTTIImpl::collectKernelLaunchBounds(), getMaxLocalMemSizeWithWaveCount(), getMaxWorkitemID(), getOccupancyWithWorkGroupSizes(), getWavesPerEU(), and makeLIDRangeMetadata().

◆ getImplicitArgNumBytes()

unsigned AMDGPUSubtarget::getImplicitArgNumBytes ( const Function & F ) const

Returns: Number of bytes of arguments that are passed to a shader or kernel in addition to the explicit ones declared for the function.

Definition at line 348 of file AMDGPUSubtarget.cpp.

References llvm::AMDGPU::AMDHSA_COV5, assert(), F, llvm::AMDGPU::getAMDHSACodeObjectVersion(), llvm::AMDGPU::isKernel(), and isMesaKernel().

Referenced by getKernArgSegmentSize().

◆ getKernArgSegmentSize()

unsigned AMDGPUSubtarget::getKernArgSegmentSize	(	const Function &	F,
		Align &	MaxAlign
	)		const

Definition at line 392 of file AMDGPUSubtarget.cpp.

References llvm::alignTo(), llvm::CallingConv::AMDGPU_KERNEL, F, getAlignmentForImplicitArgPtr(), getExplicitKernArgSize(), getExplicitKernelArgOffset(), getImplicitArgNumBytes(), and llvm::CallingConv::SPIR_KERNEL.

Referenced by llvm::AMDGPU::HSAMD::MetadataStreamerMsgPackV4::getHSAKernelProps().

◆ getLocalMemorySize()

unsigned llvm::AMDGPUSubtarget::getLocalMemorySize ( ) const

inline

Return the maximum number of bytes of LDS available for all workgroups running on the same WGP or CU.

For GFX10-GFX12 in WGP mode this is 128k even though each workgroup is limited to 64k.

Definition at line 319 of file AMDGPUSubtarget.h.

References LocalMemorySize.

Referenced by getMaxLocalMemSizeWithWaveCount(), and getOccupancyWithWorkGroupSizes().

◆ getMaxFlatWorkGroupSize()

virtual unsigned llvm::AMDGPUSubtarget::getMaxFlatWorkGroupSize ( ) const

pure virtual

Returns: Maximum flat work group size supported by the subtarget.

Implemented in llvm::GCNSubtarget, and llvm::R600Subtarget.

Referenced by getDefaultFlatWorkGroupSize(), and getFlatWorkGroupSizes().

◆ getMaxLocalMemSizeWithWaveCount()

unsigned AMDGPUSubtarget::getMaxLocalMemSizeWithWaveCount	(	unsigned	WaveCount,
		const Function &	F
	)		const

Return the amount of LDS that can be used that will not restrict the occupancy lower than WaveCount.

Definition at line 45 of file AMDGPUSubtarget.cpp.

References F, getEUsPerCU(), getFlatWorkGroupSizes(), getLocalMemorySize(), and getWavefrontSize().

◆ getMaxNumWorkGroups()

SmallVector< unsigned > AMDGPUSubtarget::getMaxNumWorkGroups ( const Function & F ) const

Return the number of work groups for the function.

Definition at line 434 of file AMDGPUSubtarget.cpp.

References F, and llvm::AMDGPU::getIntegerVecAttribute().

Referenced by llvm::GCNTTIImpl::collectKernelLaunchBounds().

◆ getMaxWavesPerEU()

unsigned llvm::AMDGPUSubtarget::getMaxWavesPerEU ( ) const

inline

Returns: Maximum number of waves per execution unit supported by the subtarget without any kind of limitation.

Definition at line 379 of file AMDGPUSubtarget.h.

References MaxWavesPerEU.

Referenced by getEffectiveWavesPerEU(), getOccupancyWithWorkGroupSizes(), and getWavesPerEU().

◆ getMaxWorkGroupsPerCU()

virtual unsigned llvm::AMDGPUSubtarget::getMaxWorkGroupsPerCU ( unsigned FlatWorkGroupSize ) const

pure virtual

Returns: Maximum number of work groups per compute unit supported by the subtarget and limited by given FlatWorkGroupSize.

Implemented in llvm::GCNSubtarget, and llvm::R600Subtarget.

Referenced by getOccupancyWithWorkGroupSizes().

◆ getMaxWorkitemID()

unsigned AMDGPUSubtarget::getMaxWorkitemID	(	const Function &	Kernel,
		unsigned	Dimension
	)		const

Return the maximum workitem ID value in the function, for the given (0, 1, 2) dimension.

Definition at line 263 of file AMDGPUSubtarget.cpp.

References getFlatWorkGroupSizes(), and getReqdWorkGroupSize().

Referenced by llvm::AMDGPUTargetLowering::computeKnownBitsForTargetNode(), isSingleLaneExecution(), llvm::AMDGPULegalizerInfo::legalizeWorkitemIDIntrinsic(), and llvm::SITargetLowering::passSpecialInputs().

◆ getMinFlatWorkGroupSize()

virtual unsigned llvm::AMDGPUSubtarget::getMinFlatWorkGroupSize ( ) const

pure virtual

Returns: Minimum flat work group size supported by the subtarget.

Implemented in llvm::GCNSubtarget, and llvm::R600Subtarget.

Referenced by getFlatWorkGroupSizes().

◆ getMinWavesPerEU()

virtual unsigned llvm::AMDGPUSubtarget::getMinWavesPerEU ( ) const

pure virtual

Returns: Minimum number of waves per execution unit supported by the subtarget.

Implemented in llvm::GCNSubtarget, and llvm::R600Subtarget.

◆ getOccupancyWithWorkGroupSizes() [1/3]

std::pair< unsigned, unsigned > AMDGPUSubtarget::getOccupancyWithWorkGroupSizes ( const MachineFunction & MF ) const

Subtarget's minimum/maximum occupancy, in number of waves per EU, that can be achieved when the only function running on a CU is MF.

This notably depends on the range of allowed flat group sizes for the function, the amount of per-workgroup LDS space required by the function, and hardware characteristics.

Definition at line 138 of file AMDGPUSubtarget.cpp.

References llvm::MachineFunction::getFunction(), llvm::MachineFunction::getInfo(), and getOccupancyWithWorkGroupSizes().

◆ getOccupancyWithWorkGroupSizes() [2/3]

std::pair< unsigned, unsigned > llvm::AMDGPUSubtarget::getOccupancyWithWorkGroupSizes	(	uint32_t	LDSBytes,
		const Function &	F
	)		const

inline

Subtarget's minimum/maximum occupancy, in number of waves per EU, that can be achieved when the only function running on a CU is F and each workgroup running the function requires LDSBytes bytes of LDS space.

This notably depends on the range of allowed flat group sizes for the function and hardware characteristics.

Definition at line 169 of file AMDGPUSubtarget.h.

References F, getFlatWorkGroupSizes(), and getOccupancyWithWorkGroupSizes().

Referenced by llvm::GCNSchedStage::checkScheduling(), llvm::GCNSubtarget::computeOccupancy(), getEffectiveWavesPerEU(), getOccupancyWithWorkGroupSizes(), and llvm::SIRegisterInfo::getRegPressureLimit().

◆ getOccupancyWithWorkGroupSizes() [3/3]

std::pair< unsigned, unsigned > AMDGPUSubtarget::getOccupancyWithWorkGroupSizes	(	uint32_t	LDSBytes,
		std::pair< unsigned, unsigned >	FlatWorkGroupSizes
	)		const

Overload which uses the specified values for the flat work group sizes, rather than querying the function itself.

FlatWorkGroupSizes should correspond to the function's value for getFlatWorkGroupSizes.

Definition at line 58 of file AMDGPUSubtarget.cpp.

References llvm::divideCeil(), getEUsPerCU(), getLocalMemorySize(), getMaxWavesPerEU(), getMaxWorkGroupsPerCU(), getWavefrontSize(), and std::swap().

◆ getReqdWorkGroupSize()

std::optional< unsigned > AMDGPUSubtarget::getReqdWorkGroupSize	(	const Function &	F,
		unsigned	Dim
	)		const

Returns: The required size of workgroups that will be used to execute F in the Dim dimension, if it is known (from !reqd_work_group_size metadata. Otherwise, returns std::nullopt.

Definition at line 233 of file AMDGPUSubtarget.cpp.

References llvm::GlobalObject::getMetadata().

Referenced by getMaxWorkitemID(), llvm::GCNTTIImpl::isSourceOfDivergence(), and makeLIDRangeMetadata().

◆ getWavefrontSize()

unsigned llvm::AMDGPUSubtarget::getWavefrontSize ( ) const

inline

◆ getWavefrontSizeLog2()

unsigned llvm::AMDGPUSubtarget::getWavefrontSizeLog2 ( ) const

inline

Definition at line 311 of file AMDGPUSubtarget.h.

References WavefrontSizeLog2.

Referenced by llvm::SIRegisterInfo::eliminateFrameIndex(), llvm::GCNSubtarget::getKnownHighZeroBitsForFrameIndex(), llvm::GCNSubtarget::initializeSubtargetDependencies(), llvm::GCNTTIImpl::isAlwaysUniform(), llvm::AMDGPULegalizerInfo::legalizeAddrSpaceCast(), llvm::SITargetLowering::LowerDYNAMIC_STACKALLOC(), and llvm::GCNTTIImpl::simplifyDemandedLaneMaskArg().

◆ getWavesPerEU() [1/3]

std::pair< unsigned, unsigned > AMDGPUSubtarget::getWavesPerEU ( const Function & F ) const

Returns: Subtarget's default pair of minimum/maximum number of waves per execution unit for function F, or minimum/maximum number of waves per execution unit explicitly requested using "amdgpu-waves-per-eu" attribute attached to function F.; Subtarget's default values if explicitly requested values cannot be converted to integer, violate subtarget's specifications, or are not compatible with minimum/maximum number of waves limited by flat work group size, register usage, and/or lds usage.

Definition at line 209 of file AMDGPUSubtarget.cpp.

References F, getFlatWorkGroupSizes(), llvm::AMDGPU::getIntegerPairAttribute(), and getWavesPerEU().

Referenced by llvm::GCNTTIImpl::collectKernelLaunchBounds(), llvm::GCNSubtarget::getMaxNumSGPRs(), llvm::GCNSubtarget::getMaxNumVGPRs(), and getWavesPerEU().

◆ getWavesPerEU() [2/3]

std::pair< unsigned, unsigned > llvm::AMDGPUSubtarget::getWavesPerEU	(	const Function &	F,
		std::pair< unsigned, unsigned >	FlatWorkGroupSizes
	)		const

Overload which uses the specified values for the flat work group sizes, rather than querying the function itself.

FlatWorkGroupSizes Should correspond to the function's value for getFlatWorkGroupSizes.

◆ getWavesPerEU() [3/3]

std::pair< unsigned, unsigned > AMDGPUSubtarget::getWavesPerEU	(	std::pair< unsigned, unsigned >	FlatWorkGroupSizes,
		unsigned	LDSBytes,
		const Function &	F
	)		const

Overload which uses the specified values for the flat workgroup sizes and LDS space rather than querying the function itself.

FlatWorkGroupSizes should correspond to the function's value for getFlatWorkGroupSizes and LDSBytes to the per-workgroup LDS allocation.

Definition at line 221 of file AMDGPUSubtarget.cpp.

References llvm::Default, F, getEffectiveWavesPerEU(), llvm::AMDGPU::getIntegerPairAttribute(), and getMaxWavesPerEU().

◆ getWavesPerEUForWorkGroup()

virtual unsigned llvm::AMDGPUSubtarget::getWavesPerEUForWorkGroup ( unsigned FlatWorkGroupSize ) const

pure virtual

Returns: Number of waves per execution unit required to support the given FlatWorkGroupSize.

Implemented in llvm::GCNSubtarget, and llvm::R600Subtarget.

Referenced by getEffectiveWavesPerEU().

◆ has16BitInsts()

bool llvm::AMDGPUSubtarget::has16BitInsts ( ) const

inline

Definition at line 212 of file AMDGPUSubtarget.h.

References Has16BitInsts.

◆ hasBF16ConversionInsts()

bool llvm::AMDGPUSubtarget::hasBF16ConversionInsts ( ) const

inline

Definition at line 229 of file AMDGPUSubtarget.h.

References HasBF16ConversionInsts.

Referenced by llvm::SITargetLowering::SITargetLowering().

◆ hasBF16PackedInsts()

bool llvm::AMDGPUSubtarget::hasBF16PackedInsts ( ) const

inline

Definition at line 233 of file AMDGPUSubtarget.h.

References HasBF16PackedInsts.

Referenced by llvm::SIInstrInfo::expandPostRAPseudo(), and llvm::SITargetLowering::SITargetLowering().

◆ hasBF16TransInsts()

bool llvm::AMDGPUSubtarget::hasBF16TransInsts ( ) const

inline

Definition at line 227 of file AMDGPUSubtarget.h.

References HasBF16TransInsts.

Referenced by llvm::SITargetLowering::SITargetLowering().

◆ hasBF8ConversionScaleInsts()

bool llvm::AMDGPUSubtarget::hasBF8ConversionScaleInsts ( ) const

inline

Definition at line 241 of file AMDGPUSubtarget.h.

References HasBF8ConversionScaleInsts.

◆ hasCvtPkF16F32Inst()

bool llvm::AMDGPUSubtarget::hasCvtPkF16F32Inst ( ) const

inline

Definition at line 253 of file AMDGPUSubtarget.h.

References HasCvtPkF16F32Inst.

Referenced by llvm::AMDGPULegalizerInfo::AMDGPULegalizerInfo(), and llvm::SITargetLowering::SITargetLowering().

◆ hasDsSrc2Insts()

bool llvm::AMDGPUSubtarget::hasDsSrc2Insts ( ) const

inline

Definition at line 263 of file AMDGPUSubtarget.h.

References HasDsSrc2Insts.

◆ hasF16BF16ToFP6BF6ConversionScaleInsts()

bool llvm::AMDGPUSubtarget::hasF16BF16ToFP6BF6ConversionScaleInsts ( ) const

inline

Definition at line 249 of file AMDGPUSubtarget.h.

References HasF16BF16ToFP6BF6ConversionScaleInsts.

◆ hasF32ToF16BF16ConversionSRInsts()

bool llvm::AMDGPUSubtarget::hasF32ToF16BF16ConversionSRInsts ( ) const

inline

Definition at line 255 of file AMDGPUSubtarget.h.

References HasF32ToF16BF16ConversionSRInsts.

◆ hasFastFMAF32()

bool llvm::AMDGPUSubtarget::hasFastFMAF32 ( ) const

inline

Definition at line 299 of file AMDGPUSubtarget.h.

References FastFMAF32.

Referenced by llvm::GCNTTIImpl::getIntrinsicInstrCost(), llvm::SITargetLowering::isFMAFasterThanFMulAndFAdd(), llvm::AMDGPULegalizerInfo::legalizeFExp(), llvm::AMDGPULegalizerInfo::legalizeFlogCommon(), llvm::AMDGPULegalizerInfo::legalizeFlogUnsafe(), llvm::AMDGPUTargetLowering::lowerFEXP(), llvm::AMDGPUTargetLowering::LowerFLOGCommon(), and llvm::AMDGPUTargetLowering::LowerFLOGUnsafe().

◆ hasFminFmaxLegacy()

bool llvm::AMDGPUSubtarget::hasFminFmaxLegacy ( ) const

inline

Definition at line 291 of file AMDGPUSubtarget.h.

References HasFminFmaxLegacy.

Referenced by llvm::AMDGPUTargetLowering::performSelectCombine().

◆ hasFP4ConversionScaleInsts()

bool llvm::AMDGPUSubtarget::hasFP4ConversionScaleInsts ( ) const

inline

Definition at line 243 of file AMDGPUSubtarget.h.

References HasFP4ConversionScaleInsts.

◆ hasFP6BF6ConversionScaleInsts()

bool llvm::AMDGPUSubtarget::hasFP6BF6ConversionScaleInsts ( ) const

inline

Definition at line 245 of file AMDGPUSubtarget.h.

References HasFP6BF6ConversionScaleInsts.

◆ hasFP8ConversionScaleInsts()

bool llvm::AMDGPUSubtarget::hasFP8ConversionScaleInsts ( ) const

inline

Definition at line 239 of file AMDGPUSubtarget.h.

References HasFP8ConversionScaleInsts.

◆ hasInv2PiInlineImm()

bool llvm::AMDGPUSubtarget::hasInv2PiInlineImm ( ) const

inline

Definition at line 287 of file AMDGPUSubtarget.h.

References HasInv2PiInlineImm.

Referenced by llvm::SITargetLowering::checkAsmConstraintValA(), llvm::SIRegisterInfo::eliminateFrameIndex(), llvm::AMDGPUTargetLowering::getConstantNegateCost(), llvm::SIInstrInfo::isInlineConstant(), llvm::SIInstrInfo::isLegalAV64PseudoImm(), and llvm::SIInstrInfo::isOperandLegal().

◆ hasMadMacF32Insts()

bool llvm::AMDGPUSubtarget::hasMadMacF32Insts ( ) const

inline

Definition at line 259 of file AMDGPUSubtarget.h.

References HasMadMacF32Insts, and isGCN().

Referenced by llvm::AMDGPULegalizerInfo::AMDGPULegalizerInfo(), llvm::GCNTTIImpl::getArithmeticInstrCost(), llvm::SITargetLowering::isFMADLegal(), llvm::SITargetLowering::isFMAFasterThanFMulAndFAdd(), llvm::AMDGPULegalizerInfo::legalizeFDIV16(), llvm::AMDGPUTargetLowering::LowerDIVREM24(), llvm::AMDGPUTargetLowering::LowerUDIVREM64(), and llvm::SITargetLowering::SITargetLowering().

◆ hasMadMixInsts()

bool llvm::AMDGPUSubtarget::hasMadMixInsts ( ) const

inline

Definition at line 235 of file AMDGPUSubtarget.h.

References HasMadMixInsts.

Referenced by llvm::SITargetLowering::isFPExtFoldable().

◆ hasMulI24()

bool llvm::AMDGPUSubtarget::hasMulI24 ( ) const

inline

Definition at line 275 of file AMDGPUSubtarget.h.

References HasMulI24.

Referenced by llvm::AMDGPUTargetLowering::performMulCombine(), llvm::AMDGPUTargetLowering::performMulhsCombine(), and llvm::AMDGPUTargetLowering::performMulLoHiCombine().

◆ hasMulU24()

bool llvm::AMDGPUSubtarget::hasMulU24 ( ) const

inline

Definition at line 279 of file AMDGPUSubtarget.h.

References HasMulU24.

Referenced by llvm::AMDGPUTargetLowering::performMulCombine(), llvm::AMDGPUTargetLowering::performMulhuCombine(), and llvm::AMDGPUTargetLowering::performMulLoHiCombine().

◆ hasSDWA()

bool llvm::AMDGPUSubtarget::hasSDWA ( ) const

inline

Definition at line 267 of file AMDGPUSubtarget.h.

References HasSDWA.

Referenced by llvm::AMDGPUTargetLowering::PerformDAGCombine(), and llvm::SIInstrInfo::verifyInstruction().

◆ hasSMulHi()

bool llvm::AMDGPUSubtarget::hasSMulHi ( ) const

inline

Definition at line 283 of file AMDGPUSubtarget.h.

References HasSMulHi.

Referenced by llvm::AMDGPURegisterBankInfo::applyMappingMAD_64_32(), llvm::AMDGPUTargetLowering::performMulhsCombine(), and llvm::AMDGPUTargetLowering::performMulhuCombine().

◆ hasTrigReducedRange()

bool llvm::AMDGPUSubtarget::hasTrigReducedRange ( ) const

inline

Definition at line 295 of file AMDGPUSubtarget.h.

References HasTrigReducedRange.

Referenced by llvm::AMDGPULegalizerInfo::legalizeSinCos().

◆ hasTrue16BitInsts()

bool llvm::AMDGPUSubtarget::hasTrue16BitInsts ( ) const

inline

Return true if the subtarget supports True16 instructions.

Definition at line 217 of file AMDGPUSubtarget.h.

References HasTrue16BitInsts.

Referenced by useRealTrue16Insts().

◆ hasVOP3PInsts()

bool llvm::AMDGPUSubtarget::hasVOP3PInsts ( ) const

inline

Definition at line 271 of file AMDGPUSubtarget.h.

References HasVOP3PInsts.

Referenced by llvm::AMDGPULegalizerInfo::AMDGPULegalizerInfo(), llvm::GCNTTIImpl::getArithmeticReductionCost(), llvm::GCNTTIImpl::getIntrinsicInstrCost(), llvm::GCNTTIImpl::getMinMaxReductionCost(), llvm::GCNTTIImpl::getShuffleCost(), llvm::AMDGPUTargetLowering::isNarrowingProfitable(), and llvm::SITargetLowering::SITargetLowering().

◆ hasWavefrontsEvenlySplittingXDim()

bool AMDGPUSubtarget::hasWavefrontsEvenlySplittingXDim	(	const Function &	F,
		bool	REquiresUniformYZ = `false`
	)		const

Returns: true if F will execute in a manner that leaves the X dimensions of the workitem ID evenly tiling wavefronts - that is, if X / wavefrontsize is uniform. This is true if either the Y and Z block dimensions are known to always be 1 or if the X dimension will always be a power of 2. If RequireUniformYZ is true, it also ensures that the Y and Z workitem IDs will be uniform (so, while a (32, 2, 1) launch with wavesize64 would ordinarily pass this test, it won't with \pRequiresUniformYZ).

This information is currently only gathered from the !reqd_work_group_size metadata on F, but this may be improved in the future.

Definition at line 241 of file AMDGPUSubtarget.cpp.

References F, getWavefrontSize(), and llvm::isPowerOf2_32().

Referenced by llvm::GCNTTIImpl::isAlwaysUniform(), and llvm::GCNTTIImpl::isSourceOfDivergence().