AMDGPUAtomicOptimizer.cpp File Reference

This pass optimizes atomic operations by using a single lane of a wavefront to perform the atomic operation, thus reducing contention on that memory location. More...

#include "AMDGPU.h"
#include "GCNSubtarget.h"
#include "llvm/Analysis/DomTreeUpdater.h"
#include "llvm/Analysis/UniformityAnalysis.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstVisitor.h"
#include "llvm/IR/IntrinsicsAMDGPU.h"
#include "llvm/InitializePasses.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"

Go to the source code of this file.

Macros
#define	DEBUG_TYPE   "amdgpu-atomic-optimizer"
	of Wavefront using llvm.cttz and performs scan using readlane & writelane intrinsics

Functions
static bool	isLegalCrossLaneType (Type *Ty)
static Value *	buildNonAtomicBinOp (IRBuilder<> &B, AtomicRMWInst::BinOp Op, Value *LHS, Value *RHS)
static Constant *	getIdentityValueForAtomicOp (Type *const Ty, AtomicRMWInst::BinOp Op)
static Value *	buildMul (IRBuilder<> &B, Value *LHS, Value *RHS)
	INITIALIZE_PASS_BEGIN (AMDGPUAtomicOptimizer, DEBUG_TYPE, "AMDGPU atomic optimizations", false, false) INITIALIZE_PASS_END(AMDGPUAtomicOptimizer

Variables
	DEBUG_TYPE
AMDGPU atomic	optimizations
AMDGPU atomic	false

Detailed Description

This pass optimizes atomic operations by using a single lane of a wavefront to perform the atomic operation, thus reducing contention on that memory location.

Atomic optimizer uses following strategies to compute scan and reduced values

1. DPP - This is the most efficient implementation for scan. DPP uses Whole Wave Mode (WWM)
2. Iterative -

Definition in file AMDGPUAtomicOptimizer.cpp.

Macro Definition Documentation

DEBUG_TYPE

#define DEBUG_TYPE   "amdgpu-atomic-optimizer"

of Wavefront using llvm.cttz and performs scan using readlane & writelane intrinsics

Definition at line 36 of file AMDGPUAtomicOptimizer.cpp.

Function Documentation

buildMul()

static Value * buildMul ( IRBuilder<> &  B, Value *  LHS, Value *  RHS  )

static

Definition at line 657 of file AMDGPUAtomicOptimizer.cpp.

References B, llvm::ConstantInt::isOne(), LHS, and RHS.

buildNonAtomicBinOp()

static Value * buildNonAtomicBinOp ( IRBuilder<> &  B, AtomicRMWInst::BinOp  Op, Value *  LHS, Value *  RHS  )

static

Definition at line 359 of file AMDGPUAtomicOptimizer.cpp.

References llvm::AtomicRMWInst::Add, llvm::AtomicRMWInst::And, B, Cond, llvm::AtomicRMWInst::FAdd, llvm::AtomicRMWInst::FMax, llvm::AtomicRMWInst::FMin, llvm::AtomicRMWInst::FSub, llvm::CmpInst::ICMP_SGT, llvm::CmpInst::ICMP_SLT, llvm::CmpInst::ICMP_UGT, llvm::CmpInst::ICMP_ULT, LHS, llvm_unreachable, llvm::AtomicRMWInst::Max, llvm::AtomicRMWInst::Min, llvm::AtomicRMWInst::Or, RHS, llvm::AtomicRMWInst::Sub, llvm::AtomicRMWInst::UMax, llvm::AtomicRMWInst::UMin, and llvm::AtomicRMWInst::Xor.

getIdentityValueForAtomicOp()

static Constant * getIdentityValueForAtomicOp ( Type *const  Ty, AtomicRMWInst::BinOp  Op  )

static

Definition at line 623 of file AMDGPUAtomicOptimizer.cpp.

References llvm::AtomicRMWInst::Add, llvm::AtomicRMWInst::And, llvm::BitWidth, llvm::CallingConv::C, llvm::AtomicRMWInst::FAdd, llvm::AtomicRMWInst::FMax, llvm::AtomicRMWInst::FMin, llvm::AtomicRMWInst::FSub, llvm::Type::getContext(), llvm::Type::getFltSemantics(), llvm::APInt::getMaxValue(), llvm::APInt::getMinValue(), llvm::APFloat::getNaN(), llvm::Type::getPrimitiveSizeInBits(), llvm::APInt::getSignedMaxValue(), llvm::APInt::getSignedMinValue(), llvm::APFloat::getZero(), llvm_unreachable, llvm::AtomicRMWInst::Max, llvm::AtomicRMWInst::Min, llvm::AtomicRMWInst::Or, llvm::AtomicRMWInst::Sub, llvm::AtomicRMWInst::UMax, llvm::AtomicRMWInst::UMin, and llvm::AtomicRMWInst::Xor.

INITIALIZE_PASS_BEGIN()

INITIALIZE_PASS_BEGIN	(	AMDGPUAtomicOptimizer	,
		DEBUG_TYPE	,
		"AMDGPU atomic optimizations"	,
		false	,
		false
	)

isLegalCrossLaneType()

static bool isLegalCrossLaneType ( Type *  Ty )

static

Definition at line 181 of file AMDGPUAtomicOptimizer.cpp.

References llvm::Type::DoubleTyID, llvm::Type::FloatTyID, llvm::Type::getIntegerBitWidth(), llvm::Type::getTypeID(), llvm::Type::IntegerTyID, and Size.

Variable Documentation

DEBUG_TYPE

DEBUG_TYPE

Definition at line 994 of file AMDGPUAtomicOptimizer.cpp.

false

AMDGPU atomic false

Definition at line 995 of file AMDGPUAtomicOptimizer.cpp.

optimizations

AMDGPU atomic optimizations

Definition at line 995 of file AMDGPUAtomicOptimizer.cpp.