LLVM: llvm::AMDGPU::CoExecInfo Struct Reference

LLVM 24.0.0git

Co-execution characteristics for a multi-cycle instruction. More...

#include "Target/AMDGPU/AMDGPUCoExecInfo.h"

Public Member Functions
	CoExecInfo ()
	Default constructor - initialize to safe defaults.
CoExecMaskT	getMask (unsigned Stage) const
	Get capability mask for a stage.
bool	canCoExec (CoExecMaskT InstMask, unsigned Stage) const
	Check if an instruction class mask can co-execute at a given stage.
std::optional< unsigned >	findNextAllowedStage (CoExecMaskT InstMask, unsigned FromStage) const
	Find next stage where the instruction class is allowed.
unsigned	getStallCycles (CoExecMaskT InstMask, unsigned Stage) const
	Return the number of cycles an instruction matching `InstMask` must wait at `Stage`.
CoExecStageType	getType (unsigned Stage) const
	Get stage type for a specific stage.

Static Public Member Functions
static CoExecStageType	getStageType (CoExecMaskT Mask)
	Get stage type from mask for display.
static CoExecInfo	build (unsigned UnitOccupancy, unsigned TotalWindow, const char *Pattern)
	Build a CoExecInfo from an occupancy and stage pattern.

Public Attributes
unsigned	UnitOccupancy = 0
	Number of cycles for which the producing instruction occupies its execution unit.
unsigned	TotalWindow = 0
	Number of cycles in the co-execution window, counting any trailing vacant stages.
CoExecSlotInfo	Slots [MaxCoExecStages]
	Per-stage slot info (capability mask).
StringRef	Pattern
	Pattern string for display (e.g., "0EIIEEIIV").

Detailed Description

Co-execution characteristics for a multi-cycle instruction.

Definition at line 237 of file AMDGPUCoExecInfo.h.

Constructor & Destructor Documentation

◆ CoExecInfo()

llvm::AMDGPU::CoExecInfo::CoExecInfo ( )

inline

Default constructor - initialize to safe defaults.

Definition at line 250 of file AMDGPUCoExecInfo.h.

References llvm::AMDGPU::All, llvm::AMDGPU::I, llvm::AMDGPU::MaxCoExecStages, and Slots.

Referenced by build().

Member Function Documentation

◆ build()

CoExecInfo llvm::AMDGPU::CoExecInfo::build	(	unsigned	UnitOccupancy,
		unsigned	TotalWindow,
		const char *	Pattern )

inlinestatic

Build a CoExecInfo from an occupancy and stage pattern.

Build CoExecInfo from a pattern string.

Pattern chars: '0'=E0, 'E'=External, 'I'=Internal, 'V'=Vacant, 'S'=Internal+ScaleWMMAAbsorb (I plus next scaled WMMA), 'T'=TRANS co-exec (all except TRANS), 'A'=Any

Definition at line 332 of file AMDGPUCoExecInfo.h.

References llvm::AMDGPU::All, assert(), CoExecInfo(), llvm::AMDGPU::I, llvm::AMDGPU::MaxCoExecStages, Pattern, llvm::AMDGPU::StageE, llvm::AMDGPU::StageE0, llvm::AMDGPU::StageI, llvm::AMDGPU::StageIS, llvm::AMDGPU::StageTR, llvm::AMDGPU::StageV, TotalWindow, and UnitOccupancy.

Referenced by llvm::AMDGPU::getCoExecInfo(), and llvm::AMDGPU::getMFMACoExecInfo().

◆ canCoExec()

bool llvm::AMDGPU::CoExecInfo::canCoExec	(	CoExecMaskT	InstMask,
		unsigned	Stage ) const

inline

Check if an instruction class mask can co-execute at a given stage.

Definition at line 261 of file AMDGPUCoExecInfo.h.

References Slots, and TotalWindow.

Referenced by getStallCycles().

◆ findNextAllowedStage()

std::optional< unsigned > llvm::AMDGPU::CoExecInfo::findNextAllowedStage	(	CoExecMaskT	InstMask,
		unsigned	FromStage ) const

inline

Find next stage where the instruction class is allowed.

Definition at line 268 of file AMDGPUCoExecInfo.h.

References llvm::AMDGPU::I, Slots, and TotalWindow.

Referenced by getStallCycles().

◆ getMask()

CoExecMaskT llvm::AMDGPU::CoExecInfo::getMask ( unsigned Stage ) const

inline

Get capability mask for a stage.

Definition at line 256 of file AMDGPUCoExecInfo.h.

References llvm::AMDGPU::All, Slots, and TotalWindow.

Referenced by getType().

◆ getStageType()

CoExecStageType llvm::AMDGPU::CoExecInfo::getStageType ( CoExecMaskT Mask )

inlinestatic

Get stage type from mask for display.

Definition at line 293 of file AMDGPUCoExecInfo.h.

References llvm::AMDGPU::E, llvm::AMDGPU::E0, llvm::AMDGPU::I, llvm::AMDGPU::IS, llvm::AMDGPU::StageE, llvm::AMDGPU::StageE0, llvm::AMDGPU::StageI, llvm::AMDGPU::StageIS, llvm::AMDGPU::StageTR, llvm::AMDGPU::StageV, llvm::AMDGPU::TR, llvm::AMDGPU::V, llvm::AMDGPU::VALU, and llvm::AMDGPU::WMMA.

Referenced by getType().

◆ getStallCycles()

unsigned llvm::AMDGPU::CoExecInfo::getStallCycles	(	CoExecMaskT	InstMask,
		unsigned	Stage ) const

inline

Return the number of cycles an instruction matching InstMask must wait at Stage.

Returns zero when no stall is required.

Definition at line 279 of file AMDGPUCoExecInfo.h.

References canCoExec(), findNextAllowedStage(), llvm::Next, and TotalWindow.

◆ getType()

CoExecStageType llvm::AMDGPU::CoExecInfo::getType ( unsigned Stage ) const

inline

Get stage type for a specific stage.

Definition at line 315 of file AMDGPUCoExecInfo.h.

References getMask(), and getStageType().

Member Data Documentation

◆ Pattern

StringRef llvm::AMDGPU::CoExecInfo::Pattern

Pattern string for display (e.g., "0EIIEEIIV").

Definition at line 247 of file AMDGPUCoExecInfo.h.

Referenced by build(), and llvm::AMDGPU::getMFMACoExecInfo().

◆ Slots

CoExecSlotInfo llvm::AMDGPU::CoExecInfo::Slots[MaxCoExecStages]

Per-stage slot info (capability mask).

Definition at line 245 of file AMDGPUCoExecInfo.h.

Referenced by canCoExec(), CoExecInfo(), findNextAllowedStage(), getMask(), and llvm::AMDGPU::getMFMACoExecInfo().

◆ TotalWindow

unsigned llvm::AMDGPU::CoExecInfo::TotalWindow = 0

Number of cycles in the co-execution window, counting any trailing vacant stages.

Definition at line 243 of file AMDGPUCoExecInfo.h.

Referenced by build(), canCoExec(), findNextAllowedStage(), getMask(), llvm::AMDGPU::getMFMACoExecInfo(), and getStallCycles().

◆ UnitOccupancy

unsigned llvm::AMDGPU::CoExecInfo::UnitOccupancy = 0

Number of cycles for which the producing instruction occupies its execution unit.

Definition at line 240 of file AMDGPUCoExecInfo.h.

Referenced by build().

The documentation for this struct was generated from the following file:

lib/Target/AMDGPU/AMDGPUCoExecInfo.h