llvm-project/llvm/lib/Target/AMDGPU/AMDGPUAtomicOptimizer.cpp

//===-- AMDGPUAtomicOptimizer.cpp -----------------------------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file
/// 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.
//
//===----------------------------------------------------------------------===//

#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "llvm/Analysis/LegacyDivergenceAnalysis.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstVisitor.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"

#define DEBUG_TYPE "amdgpu-atomic-optimizer"

using namespace llvm;

namespace {

enum DPP_CTRL {
  DPP_ROW_SR1 = 0x111,
  DPP_ROW_SR2 = 0x112,
  DPP_ROW_SR4 = 0x114,
  DPP_ROW_SR8 = 0x118,
  DPP_WF_SR1 = 0x138,
  DPP_ROW_BCAST15 = 0x142,
  DPP_ROW_BCAST31 = 0x143
};

struct ReplacementInfo {
  Instruction *I;
  Instruction::BinaryOps Op;
  unsigned ValIdx;
  bool ValDivergent;
};

class AMDGPUAtomicOptimizer : public FunctionPass,
                              public InstVisitor<AMDGPUAtomicOptimizer> {
private:
  SmallVector<ReplacementInfo, 8> ToReplace;
  const LegacyDivergenceAnalysis *DA;
  const DataLayout *DL;
  DominatorTree *DT;
  bool HasDPP;

  void optimizeAtomic(Instruction &I, Instruction::BinaryOps Op,
                      unsigned ValIdx, bool ValDivergent) const;

  void setConvergent(CallInst *const CI) const;

public:
  static char ID;

  AMDGPUAtomicOptimizer() : FunctionPass(ID) {}

  bool runOnFunction(Function &F) override;

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addPreserved<DominatorTreeWrapperPass>();
    AU.addRequired<LegacyDivergenceAnalysis>();
    AU.addRequired<TargetPassConfig>();
  }

  void visitAtomicRMWInst(AtomicRMWInst &I);
  void visitIntrinsicInst(IntrinsicInst &I);
};

} // namespace

char AMDGPUAtomicOptimizer::ID = 0;

char &llvm::AMDGPUAtomicOptimizerID = AMDGPUAtomicOptimizer::ID;

bool AMDGPUAtomicOptimizer::runOnFunction(Function &F) {
  if (skipFunction(F)) {
    return false;
  }

  DA = &getAnalysis<LegacyDivergenceAnalysis>();
  DL = &F.getParent()->getDataLayout();
  DominatorTreeWrapperPass *const DTW =
      getAnalysisIfAvailable<DominatorTreeWrapperPass>();
  DT = DTW ? &DTW->getDomTree() : nullptr;
  const TargetPassConfig &TPC = getAnalysis<TargetPassConfig>();
  const TargetMachine &TM = TPC.getTM<TargetMachine>();
  const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
  HasDPP = ST.hasDPP();

  visit(F);

  const bool Changed = !ToReplace.empty();

  for (ReplacementInfo &Info : ToReplace) {
    optimizeAtomic(*Info.I, Info.Op, Info.ValIdx, Info.ValDivergent);
  }

  ToReplace.clear();

  return Changed;
}

void AMDGPUAtomicOptimizer::visitAtomicRMWInst(AtomicRMWInst &I) {
  // Early exit for unhandled address space atomic instructions.
  switch (I.getPointerAddressSpace()) {
  default:
    return;
  case AMDGPUAS::GLOBAL_ADDRESS:
  case AMDGPUAS::LOCAL_ADDRESS:
    break;
  }

  Instruction::BinaryOps Op;

  switch (I.getOperation()) {
  default:
    return;
  case AtomicRMWInst::Add:
    Op = Instruction::Add;
    break;
  case AtomicRMWInst::Sub:
    Op = Instruction::Sub;
    break;
  }

  const unsigned PtrIdx = 0;
  const unsigned ValIdx = 1;

  // If the pointer operand is divergent, then each lane is doing an atomic
  // operation on a different address, and we cannot optimize that.
  if (DA->isDivergent(I.getOperand(PtrIdx))) {
    return;
  }

  const bool ValDivergent = DA->isDivergent(I.getOperand(ValIdx));

  // If the value operand is divergent, each lane is contributing a different
  // value to the atomic calculation. We can only optimize divergent values if
  // we have DPP available on our subtarget, and the atomic operation is 32
  // bits.
  if (ValDivergent && (!HasDPP || (DL->getTypeSizeInBits(I.getType()) != 32))) {
    return;
  }

  // If we get here, we can optimize the atomic using a single wavefront-wide
  // atomic operation to do the calculation for the entire wavefront, so
  // remember the instruction so we can come back to it.
  const ReplacementInfo Info = {&I, Op, ValIdx, ValDivergent};

  ToReplace.push_back(Info);
}

void AMDGPUAtomicOptimizer::visitIntrinsicInst(IntrinsicInst &I) {
  Instruction::BinaryOps Op;

  switch (I.getIntrinsicID()) {
  default:
    return;
  case Intrinsic::amdgcn_buffer_atomic_add:
  case Intrinsic::amdgcn_struct_buffer_atomic_add:
  case Intrinsic::amdgcn_raw_buffer_atomic_add:
    Op = Instruction::Add;
    break;
  case Intrinsic::amdgcn_buffer_atomic_sub:
  case Intrinsic::amdgcn_struct_buffer_atomic_sub:
  case Intrinsic::amdgcn_raw_buffer_atomic_sub:
    Op = Instruction::Sub;
    break;
  }

  const unsigned ValIdx = 0;

  const bool ValDivergent = DA->isDivergent(I.getOperand(ValIdx));

  // If the value operand is divergent, each lane is contributing a different
  // value to the atomic calculation. We can only optimize divergent values if
  // we have DPP available on our subtarget, and the atomic operation is 32
  // bits.
  if (ValDivergent && (!HasDPP || (DL->getTypeSizeInBits(I.getType()) != 32))) {
    return;
  }

  // If any of the other arguments to the intrinsic are divergent, we can't
  // optimize the operation.
  for (unsigned Idx = 1; Idx < I.getNumOperands(); Idx++) {
    if (DA->isDivergent(I.getOperand(Idx))) {
      return;
    }
  }

  // If we get here, we can optimize the atomic using a single wavefront-wide
  // atomic operation to do the calculation for the entire wavefront, so
  // remember the instruction so we can come back to it.
  const ReplacementInfo Info = {&I, Op, ValIdx, ValDivergent};

  ToReplace.push_back(Info);
}

void AMDGPUAtomicOptimizer::optimizeAtomic(Instruction &I,
                                           Instruction::BinaryOps Op,
                                           unsigned ValIdx,
                                           bool ValDivergent) const {
  LLVMContext &Context = I.getContext();

  // Start building just before the instruction.
  IRBuilder<> B(&I);

  Type *const Ty = I.getType();
  const unsigned TyBitWidth = DL->getTypeSizeInBits(Ty);
  Type *const VecTy = VectorType::get(B.getInt32Ty(), 2);

  // This is the value in the atomic operation we need to combine in order to
  // reduce the number of atomic operations.
  Value *const V = I.getOperand(ValIdx);

  // We need to know how many lanes are active within the wavefront, and we do
  // this by getting the exec register, which tells us all the lanes that are
  // active.
  MDNode *const RegName =
      llvm::MDNode::get(Context, llvm::MDString::get(Context, "exec"));
  Value *const Metadata = llvm::MetadataAsValue::get(Context, RegName);
  CallInst *const Exec =
      B.CreateIntrinsic(Intrinsic::read_register, {B.getInt64Ty()}, {Metadata});
  setConvergent(Exec);

  // We need to know how many lanes are active within the wavefront that are
  // below us. If we counted each lane linearly starting from 0, a lane is
  // below us only if its associated index was less than ours. We do this by
  // using the mbcnt intrinsic.
  Value *const BitCast = B.CreateBitCast(Exec, VecTy);
  Value *const ExtractLo = B.CreateExtractElement(BitCast, B.getInt32(0));
  Value *const ExtractHi = B.CreateExtractElement(BitCast, B.getInt32(1));
  CallInst *const PartialMbcnt = B.CreateIntrinsic(
      Intrinsic::amdgcn_mbcnt_lo, {}, {ExtractLo, B.getInt32(0)});
  CallInst *const Mbcnt = B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_hi, {},
                                            {ExtractHi, PartialMbcnt});

  Value *const MbcntCast = B.CreateIntCast(Mbcnt, Ty, false);

  Value *LaneOffset = nullptr;
  Value *NewV = nullptr;

  // If we have a divergent value in each lane, we need to combine the value
  // using DPP.
  if (ValDivergent) {
    // First we need to set all inactive invocations to 0, so that they can
    // correctly contribute to the final result.
    CallInst *const SetInactive = B.CreateIntrinsic(
        Intrinsic::amdgcn_set_inactive, Ty, {V, B.getIntN(TyBitWidth, 0)});
    setConvergent(SetInactive);
    NewV = SetInactive;

    const unsigned Iters = 6;
    const unsigned DPPCtrl[Iters] = {DPP_ROW_SR1,     DPP_ROW_SR2,
                                     DPP_ROW_SR4,     DPP_ROW_SR8,
                                     DPP_ROW_BCAST15, DPP_ROW_BCAST31};
    const unsigned RowMask[Iters] = {0xf, 0xf, 0xf, 0xf, 0xa, 0xc};

    // This loop performs an inclusive scan across the wavefront, with all lanes
    // active (by using the WWM intrinsic).
    for (unsigned Idx = 0; Idx < Iters; Idx++) {
      CallInst *const DPP = B.CreateIntrinsic(Intrinsic::amdgcn_mov_dpp, Ty,
                                              {NewV, B.getInt32(DPPCtrl[Idx]),
                                               B.getInt32(RowMask[Idx]),
                                               B.getInt32(0xf), B.getFalse()});
      setConvergent(DPP);
      Value *const WWM = B.CreateIntrinsic(Intrinsic::amdgcn_wwm, Ty, DPP);

      NewV = B.CreateBinOp(Op, NewV, WWM);
      NewV = B.CreateIntrinsic(Intrinsic::amdgcn_wwm, Ty, NewV);
    }

    // NewV has returned the inclusive scan of V, but for the lane offset we
    // require an exclusive scan. We do this by shifting the values from the
    // entire wavefront right by 1, and by setting the bound_ctrl (last argument
    // to the intrinsic below) to true, we can guarantee that 0 will be shifted
    // into the 0'th invocation.
    CallInst *const DPP =
        B.CreateIntrinsic(Intrinsic::amdgcn_mov_dpp, {Ty},
                          {NewV, B.getInt32(DPP_WF_SR1), B.getInt32(0xf),
                           B.getInt32(0xf), B.getTrue()});
    setConvergent(DPP);
    LaneOffset = B.CreateIntrinsic(Intrinsic::amdgcn_wwm, Ty, DPP);

    // Read the value from the last lane, which has accumlated the values of
    // each active lane in the wavefront. This will be our new value with which
    // we will provide to the atomic operation.
    if (TyBitWidth == 64) {
      Value *const ExtractLo = B.CreateTrunc(NewV, B.getInt32Ty());
      Value *const ExtractHi =
          B.CreateTrunc(B.CreateLShr(NewV, B.getInt64(32)), B.getInt32Ty());
      CallInst *const ReadLaneLo = B.CreateIntrinsic(
          Intrinsic::amdgcn_readlane, {}, {ExtractLo, B.getInt32(63)});
      setConvergent(ReadLaneLo);
      CallInst *const ReadLaneHi = B.CreateIntrinsic(
          Intrinsic::amdgcn_readlane, {}, {ExtractHi, B.getInt32(63)});
      setConvergent(ReadLaneHi);
      Value *const PartialInsert = B.CreateInsertElement(
          UndefValue::get(VecTy), ReadLaneLo, B.getInt32(0));
      Value *const Insert =
          B.CreateInsertElement(PartialInsert, ReadLaneHi, B.getInt32(1));
      NewV = B.CreateBitCast(Insert, Ty);
    } else if (TyBitWidth == 32) {
      CallInst *const ReadLane = B.CreateIntrinsic(Intrinsic::amdgcn_readlane,
                                                   {}, {NewV, B.getInt32(63)});
      setConvergent(ReadLane);
      NewV = ReadLane;
    } else {
      llvm_unreachable("Unhandled atomic bit width");
    }
  } else {
    // Get the total number of active lanes we have by using popcount.
    Instruction *const Ctpop = B.CreateUnaryIntrinsic(Intrinsic::ctpop, Exec);
    Value *const CtpopCast = B.CreateIntCast(Ctpop, Ty, false);

    // Calculate the new value we will be contributing to the atomic operation
    // for the entire wavefront.
    NewV = B.CreateMul(V, CtpopCast);
    LaneOffset = B.CreateMul(V, MbcntCast);
  }

  // We only want a single lane to enter our new control flow, and we do this
  // by checking if there are any active lanes below us. Only one lane will
  // have 0 active lanes below us, so that will be the only one to progress.
  Value *const Cond = B.CreateICmpEQ(MbcntCast, B.getIntN(TyBitWidth, 0));

  // Store I's original basic block before we split the block.
  BasicBlock *const EntryBB = I.getParent();

  // We need to introduce some new control flow to force a single lane to be
  // active. We do this by splitting I's basic block at I, and introducing the
  // new block such that:
  // entry --> single_lane -\
  //       \------------------> exit
  Instruction *const SingleLaneTerminator =
      SplitBlockAndInsertIfThen(Cond, &I, false, nullptr, DT, nullptr);

  // Move the IR builder into single_lane next.
  B.SetInsertPoint(SingleLaneTerminator);

  // Clone the original atomic operation into single lane, replacing the
  // original value with our newly created one.
  Instruction *const NewI = I.clone();
  B.Insert(NewI);
  NewI->setOperand(ValIdx, NewV);

  // Move the IR builder into exit next, and start inserting just before the
  // original instruction.
  B.SetInsertPoint(&I);

  // Create a PHI node to get our new atomic result into the exit block.
  PHINode *const PHI = B.CreatePHI(Ty, 2);
  PHI->addIncoming(UndefValue::get(Ty), EntryBB);
  PHI->addIncoming(NewI, SingleLaneTerminator->getParent());

  // We need to broadcast the value who was the lowest active lane (the first
  // lane) to all other lanes in the wavefront. We use an intrinsic for this,
  // but have to handle 64-bit broadcasts with two calls to this intrinsic.
  Value *BroadcastI = nullptr;

  if (TyBitWidth == 64) {
    Value *const ExtractLo = B.CreateTrunc(PHI, B.getInt32Ty());
    Value *const ExtractHi =
        B.CreateTrunc(B.CreateLShr(PHI, B.getInt64(32)), B.getInt32Ty());
    CallInst *const ReadFirstLaneLo =
        B.CreateIntrinsic(Intrinsic::amdgcn_readfirstlane, {}, ExtractLo);
    setConvergent(ReadFirstLaneLo);
    CallInst *const ReadFirstLaneHi =
        B.CreateIntrinsic(Intrinsic::amdgcn_readfirstlane, {}, ExtractHi);
    setConvergent(ReadFirstLaneHi);
    Value *const PartialInsert = B.CreateInsertElement(
        UndefValue::get(VecTy), ReadFirstLaneLo, B.getInt32(0));
    Value *const Insert =
        B.CreateInsertElement(PartialInsert, ReadFirstLaneHi, B.getInt32(1));
    BroadcastI = B.CreateBitCast(Insert, Ty);
  } else if (TyBitWidth == 32) {
    CallInst *const ReadFirstLane =
        B.CreateIntrinsic(Intrinsic::amdgcn_readfirstlane, {}, PHI);
    setConvergent(ReadFirstLane);
    BroadcastI = ReadFirstLane;
  } else {
    llvm_unreachable("Unhandled atomic bit width");
  }

  // Now that we have the result of our single atomic operation, we need to
  // get our individual lane's slice into the result. We use the lane offset we
  // previously calculated combined with the atomic result value we got from the
  // first lane, to get our lane's index into the atomic result.
  Value *const Result = B.CreateBinOp(Op, BroadcastI, LaneOffset);

  // Replace the original atomic instruction with the new one.
  I.replaceAllUsesWith(Result);

  // And delete the original.
  I.eraseFromParent();
}

void AMDGPUAtomicOptimizer::setConvergent(CallInst *const CI) const {
  CI->addAttribute(AttributeList::FunctionIndex, Attribute::Convergent);
}

INITIALIZE_PASS_BEGIN(AMDGPUAtomicOptimizer, DEBUG_TYPE,
                      "AMDGPU atomic optimizations", false, false)
INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis)
INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
INITIALIZE_PASS_END(AMDGPUAtomicOptimizer, DEBUG_TYPE,
                    "AMDGPU atomic optimizations", false, false)

FunctionPass *llvm::createAMDGPUAtomicOptimizerPass() {
  return new AMDGPUAtomicOptimizer();
}